Thèse de Doctorat: Parth Raj S

Thèse de Doctorat
Parth Raj S INGH

Mémoire présenté en vue de l’obtention du
grade de Docteur de l’Université de Nantes
sous le sceau de l’Université Bretagne Loire
École doctorale : Sciences et technologies de l’information, et mathématiques (STIM-503)
Discipline : Électronique
Spécialité : Communications numériques
Unité de recherche : IETR UMR CNRS 6164
Soutenue le 12 October 2017
Source Localization with MIMO Systems
JURY
Président : M. Jean-Pierre C ANCES, Professeur, Université de Limoges
Rapporteurs : M. Rémy B OYER, Maître de conférences/HDR, CentraleSupélec
M. Salah B OURENNANE, Professeur, École Centrale Marseille
Directeur de thèse : M. Yide WANG, Professeur, Université de Nantes
Co-directeur de thèse : M. Pascal C HARGÉ, Professeur, Université de Nantes
Remerciements
During this three-year research work at Institute of Electronics and Telecommunica-

tion of Rennes (IETR), Polytech Nantes, I have been academically and psychologically
supported by many people to whom I am very grateful.
First of all, I would like to convey my sincere gratitude to the director of my thesis,
Professor Yide Wang, for offering me the opportunity to pursue PhD (Master as well)
under his supervision. I am also very grateful to the co-director of my thesis, Professor
Pascal Chargé for his useful remarks and encouragement during this period. I am
very thankful for their precious time, motivation, expert guidance, understanding, and
patience for the last four years. Without their well-designed research plan and clear
goals, I couldn’t have completed my PhD.
I would like to express my appreciation to the committee members, Professor Jean-
Pierre Cances, Professor Salah Bourennane, and Professor Rémy Boyer, for accepting
the request to validate my thesis and participate in my oral defense. I am also thankful
for their insightful comments and positive suggestions.
I would like to thank Mrs. Sandrine Charlier for establishing a solid foundation
for our academic life and handling all the administrative affairs, which has greatly
facilitated my research. I would also like to thank Mr. Marc Brunet and Mr. Guillaume
Lirzin for their technical support and activeness in organizing social events. Along
with them, IETR is blessed with many professors like Professor Jean-François Diouris,
Professor Tchanguiz Razban, Professor Christophe Bourlier, Professor Safwan El Assad
and so on who are always ready to provide academic assistance.
My colleagues have also played a significant role during my PhD by sharing their
knowledge and experience. They are Dušan Cvetković, Mohammed Abutaha, Nabil
Zaraneh, Christophe Bouguet, Ons Jallouli, Ali Mohydeen, Jingjing Pan, Meng Sun,
Randa Jaouadi, Taoufik Bouguera, Simei Yang, Jinze Du, Irfan Ali Tunio, Yunniel Aren-
cibia Noa, Julien Weckbrodt and few more. Specially, I would like to express additional
gratitude to Jingjing Pan and Ali Mohydeen for their useful academic discussions.
Moreover, I feel honored to have Amit Karel, Bhanu Pratap Singh Jadav, and Shub-
ham Maheshwari as my best friends for their unconditional friendship and moral sup-
port during my PhD. A special thank also goes to Marion Giraud, Mr. Jean Marie
Barth, Erwan Quemere, Armand Bax, and Armand Fuchs for introducing me to the
3
4 Remerciements
French culture and cuisine. They have greatly enriched my life in France.
Finally, from the bottom of my heart I would like to thank my mother Heeramani,
my father Nanku Ram, and my little sister Divya Shail for their endless love, support
and encouragement throughout my endeavors.
Résumé
La localisation des cibles est présente dans notre vie quotidienne. La plupart des
animaux et les humains sont équipés de certains types de systèmes de localisation. Le
fait d’avoir deux oreilles est la première preuve. Avoir deux oreilles aide les animaux à
localiser la direction de la source sonore. En outre, les chauves-souris utilisent des on-
des ultrasonores pour naviguer dans l’obscurité. Inspiré de la nature, le traitement des
signaux électromagnétiques et acoustiques a été étudié pendant de nombreuses décen-
nies pour localiser les sources radio et acoustiques, respectivement. La localisation de
sources est utilisée dans le radar, le sonar et même les télécommunications. Cepen-
dant, grâce à la capacité à se propager dans l’espace, à grande vitesse et à une petite
longueur d’onde des ondes radio, les radars ont de nombreux avantages par rapport
au sonar, tels qu’une résolution élevée, une réponse rapide, et une capacité de localiser
les sources aussi bien à l’intérieur d’une structure que dans l’espace.
Le radar a attiré l’attention pendant la Seconde Guerre mondiale lorsqu’il a sur-
passé les capacités visuelles des humains pour détecter les avions et les navires. Le
radar a de nombreuses applications civiles et militaires qui nous permettent de l’utiliser
directement ou indirectement. Il joue un rôle dominant dans la navigation des trans-
ports comme les avions, les bateaux, les navires et les voitures intelligentes. Les radars
sont également utilisés dans les prévisions météorologiques, le suivi de météorites, la
surveillance pour des raisons de sécurité, les contrôles de la circulation, les recherches
scientifiques et bien d’autres domaines. Un radar est distingué en fonction de ses ap-
plications, du type de forme d’onde utilisé et de ses principales caractéristiques. En
conséquence, il existe de nombreuses façons de classer un système radar. Une de ces
classifications consiste à distinguer les radars passifs et actifs. Un système radar pas-
sif est composé uniquement d’antennes réceptrices. Il s’appuie sur les signaux émis
par les sources ou par une antenne de transmission non-coopérative qui illumine les
sources. D’autre part, le système radar actif contient des éléments transmetteurs et ré-
cepteurs. En raison des antennes de transmission dédiées, nous connaissons les formes
d’onde émises qui nous permettent d’extraire plus d’informations sur les cibles des
échos reçus. Initialement, des antennes directives ont été montées sur une configura-
tion mécanique rotative pour scruter la région d’intérêt. Plus tard, cette configuration
est remplacée par un réseau de déphaseurs qui peut diriger le faisceau à grande vitesse
5
6 Résumé
dans la région d’intérêt. Nous pouvons trouver une énorme quantité de travail sur les
techniques de la localisation de sources en utilisant des radars passifs et actifs avec
des antennes multiples. La formation de voies, MVDR, MUSIC, ESPRIT, etc. sont des
méthodes d’estimation de DOA les plus célèbres.
Depuis le début, le radar a connu de nombreuses évolutions. Récemment, la tech-
nique de communication basée sur MIMO a inspiré le développement du radar MIMO.
L’amélioration des composantes hyperfréquences et des architectures de processeurs
rapides a également joué un rôle très important dans le développement d’un radar
MIMO. Un radar MIMO se compose de plusieurs antennes d’émission et de récep-
tion comme un radar à balayage de faisceaux classique. Cependant, les antennes
d’émission dans un radar MIMO transmettent des signaux linéairement indépendants
qui peuvent être facilement identifiés par les filtres adaptés au niveau du récepteur.
Lors de l’interaction des signaux orthogonaux transmis avec les cibles, chaque signal
transmis possède individuellement des informations sur les cibles. Après les filtrages
adaptés des signaux reçus, les informations sur les cibles sont traitées afin d’extraire
les distances, les DOAs, les DODs, les vitesses, les tailles, etc. de ces cibles. Le radar
MIMO fournit une haute résolution angulaire, une grande diversité spatiale et une ex-
cellente identifiabilité des paramètres. Dans un sens large, un système de radar MIMO
peut être classé comme statistique ou cohérent. Un radar MIMO statistique a des an-
tennes largement séparées alors qu’un radar MIMO cohérent a plutôt des antennes
co-localisées. Cette thèse est consacrée aux systèmes radar MIMO avec des antennes
co-localisées, précisément, un système MIMO bistatique dont les réseaux d’émission
et de réception avec antennes co-localisées sont séparés d’une grande distance.
En fonction de la distance d’une cible par rapport au centre du réseau d’antennes,
la cible peut être considérée comme située en champ proche ou en champ lointain de
ce réseau. En champ lointain, le front d’onde sphérique d’une onde est approximé
comme planaire. La plupart des méthodes de localisation de sources existantes sont
consacrées à la situation du champ lointain. Cependant, la localisation de cibles en
champ proche est également importante en raison de ses nombreuses applications,
notamment à l’intérieur des bâtiments. Dans la zone du champ proche d’un réseau, le
front d’onde sphérique d’un signal source ne peut plus être considéré comme planaire.
Afin de réduire la non-linéarité du modèle exact du signal basé sur le front d’onde
sphérique et de simplifier les algorithmes de localisation, le front d’onde sphérique est
souvent approximé en quadrique (surface quadratique) en utilisant les termes de deux
premiers ordres de l’expansion de Taylor (appelé aussi approximation de Fresnel).
Dans cette thèse, nous avons proposé une nouvelle méthode pour localiser des
cibles en champ proche d’un système MIMO bistatique en utilisant l’approximation
de Fresnel. Dans le cas du MIMO bistatique en champ proche, nous devons estimer
quatre paramètres pour chaque cible, à savoir, la direction du départ par rapport au
Résumé 7
réseau d’antennes d’émission (DOD) ; la direction d’arrivée par rapport au réseau

d’antennes de réception (DOA) ; la distance entre la cible et le point de référence du
réseau d’émission et la distance entre la cible et le point de référence du réseau de ré-
ception. Dans cette méthode, qui est basée sur le modèle approximé des signaux reçus,
nous estimons d’abord les DODs et les DOAs en utilisant la méthode 1D RARE. En-
suite, la distance entre la cible et le réseau d’antenne d’émission est estimée en utilisant
les DODs estimés auparavant. Enfin, à partir de la distance cible-réseau d’émission et
DOD, nous estimons la distance cible-réseau de réception et en même temps que nous
réalisons l’appariement des DOAs estimés en utilisant la technique 2D MUSIC. Cette
méthode proposée est comparée à une méthode existante basée sur les sous-matrices
de la matrice de covariance. À partir de la comparaison des résultats de simulation,
on peut observer que la méthode proposée surpasse largement la méthode existante
en termes de RMSE et du nombre maximal de cibles localisables. Il convient de sig-
naler que, la méthode proposée et la méthode existante utilisent toutes les deux, des
antennes linéaires uniformes symétriques ayant une distance inter-élément égale au
quart de longueur d’onde.
Comme mentionné précédemment, dans l’approximation de Fresnel, le vrai front
d’onde sphérique est approximé comme front d’onde quadrique. Cette approximation
introduit une erreur dans le modèle du signal reçu résultant une erreur systématique
sur les paramètres de localisation estimés (biais d’estimation). Ces biais dégradent la
précision des techniques d’estimation. Il existe deux approches permettant de réduire,
voire éliminer ces bais d’estimation dus à l’approximation de Fresnel. La première
méthode consiste à réduire l’effet de l’approximation de Fresnel sur les paramètres de
localisation estimés en utilisant une méthode de correction. La deuxième approche est
d’éviter l’utilisation de l’approximation de Fresnel et de travailler directement avec le
modèle exact des signaux reçus qui est basé sur le front d’onde sphérique.
La première approche vise à réduire l’effet de l’approximation de Fresnel. Nous
avons proposé deux nouvelles méthodes pour corriger les paramètres de localisation
estimés par des méthodes basées sur le modèle approximé.
• La première méthode de correction utilise deux LUTs correspondant aux réseaux

d’émission et de réception respectivement. Les deux LUTs représentent la corre-
spondance entre les vrais paramètres de localisation et les paramètres biaisés cal-
culés à l’aide du modèle approximé de signal reçu. La région du champ proche
étant délimitée par les limites inférieure et supérieure de la région de Fresnel,
pour créer un LUT correspondant à l’émission et à la réception, la région du
champ proche est d’abord discrétisée. Ensuite, pour chaque position discrétisée
(donc les vrais paramètres de localisation), les paramètres de localisation biaisés
correspondants sont calculés en utilisant une méthode des moindres carrés pour
résoudre un système surdéterminé d’équations linéaires. Il peut y avoir plusieurs
8 Résumé
méthodes permettant d’obtenir les paramètres de localisation biaisés. Chaque

méthode utilise une approche différente qui mène une erreur différente pour les
mêmes vrais paramètres de localisation. Ainsi, nous choisissons la méthode la
moins dépendante de l’angle directionnel. Parce que l’erreur d’approximation de
Fresnel devrait idéalement diminuer lorsque la cible s’éloigne du réseau d’antennes,
il ne devrait pas avoir de dépendance critique sur l’angle directionnel. Les paramètres
de localisation biaisés peuvent être obtenus en utilisant une méthode basée sur
le modèle approximé appliquée aux signaux générés avec un modèle exacte. Les
LUT étant discrets, les paramètres de localisation estimés ne correspondent pas
nécessairement à une des entrées de LUT. Par conséquent, nous cherchons les
quatre entrées de LUT les plus proches des paramètres estimés et interpolons
en utilisant une technique d’interpolation de faible complexité pour obtenir les
paramètres de localisation corrigés. Deux LUTs sont construits, l’un servant à la
correction des paramètres liés au réseau d’antennes d’émission et l’autre pour la
réception.
• La méthode de correction basée sur LUT peut prendre un peu de temps car
elle nécessite de chercher les quatre paramètres de localisation biaisés les plus
proches des paramètres estimés. Ainsi, nous proposons une autre méthode de
correction qui estime directement les longueurs de chemin relatives à partir des
paramètres biaisés de localisation estimés en utilisant le modèle de signal ap-
proché. Une longueur de parcours relative d’une antenne d’un réseau est sim-
plement la différence entre la distance d’une cible à cette antenne et celle de
la même cible à l’antenne de référence de ce réseau. Cette méthode de correc-
tion utilise le fait que les longueurs relatives d’une cible sont mieux estimées que
les paramètres (distances et angles) par les méthodes basées sur un modèle ap-
proximé. En d’autres termes, si une méthode basée sur le modèle approximé
est appliquée sur un signal en champ proche, elle va générer des erreurs sur les
paramètres DOA, DOD et les distances, mais les erreurs sur les longueurs rela-
tives (ou bien sur le vrai vecteur directionnel) sont beaucoup plus petites. Une
fois que nous avons des estimations des longueurs relatives des chemins, nous
pouvons obtenir un système surdéterminé d’équations linéaires en utilisant le
modèle exacte basé sur le front d’onde sphérique. Nous pouvons le résoudre en
utilisant l’estimateur des moindres carrés pour obtenir les paramètres de locali-
sation corrigés.
La comparaison en termes des RMSE des paramètres de localisation avant et après
la correction montre que les méthodes de correction peuvent réduire considérablement
le biais pour des SNR élevés où les erreurs dominantes sont celles liées à l’approximation
du modèle. Les performances de correction des méthodes proposées dépendent égale-
ment de la performance de la méthode basée sur le modèle approximé utilisée pour
Résumé 9
estimer les paramètres de localisation.

En plus des deux méthodes de correction, nous avons aussi proposé deux nouvelles
méthodes qui peuvent traiter directement le modèle exacte des signaux basé sur le
front d’onde sphérique.
• La première méthode proposée basée sur le modèle exacte s’inspire d’une méth-
ode basée sur le modèle approximé existante pour localiser les cibles en champ
proche d’un radar MIMO bistatique. L’idée principale de travailler avec le mod-
èle exact du signal, est d’estimer d’abord directement les vecteurs directionnels,
puis d’en extraire les paramètres de localisation. Par conséquent, la méthode pro-
posée basée sur le modèle exacte utilise une technique s’appuyant sur les sous-
espaces pour estimer les vecteurs directionnels. À partir des vecteurs direction-
nels estimés, nous pouvons alors extraire les longueurs relatives des chemins.
Elles peuvent être utilisées pour obtenir un système surdéterminé d’équations
linéaires qui peut être résolu par un estimateur des moindres carrés pour obtenir
les paramètres de localisation. L’extraction des longueurs de chemin relatives à
partir du vecteur directionnel nécessite un déroulement (phase unwrapping) et
une correction du décalage de phase. Le déroulement de phase est nécessaire car
les composants des vecteurs directionnels sont complexes et leurs phases ne peu-
vent être récupérées que dans leur domaine principal, tandis que les longueurs
des chemins relatives sont directement proportionnelles aux phases réelles des
composants complexes des vecteurs directionnels. Le déroulement fournit habituelle-
ment des phases décalées qui peuvent être corrigées à partir du fait que la phase
de la composante de référence devrait être nulle. Ainsi, en soustrayant la phase
du composant de référence des phases restantes, on obtient les estimations des
longueurs de trajets relatives. Les longueurs de parcours estimées peuvent être
utilisées pour estimer les paramètres de localisation en utilisant la procédure util-
isée par la deuxième méthode de correction proposée.
• La deuxième méthode basée sur le modèle exacte utilise la décomposition des

tenseurs pour estimer les vecteurs directionnels. Les signaux reçus d’un système
MIMO bistatique peuvent être représentés pas un tenseur d’ordre trois qui peut
être décomposé en vecteurs directionnels par PARAFAC. Une fois que les esti-
mations des vecteurs directionnels sont disponibles, la même approche, qui est
utilisée dans la méthode proposé ci-dessus, peut être utilisée pour obtenir les
paramètres de localisation.
La comparaison entre les deux méthodes proposées basées sur le modèle exacte
révèle que la méthode basée sur PARAFAC a des performances légèrement meilleures,
notamment parce que la première méthode n’utilise pas la totalité de l’information
disponible. Les méthodes basées sur le modèle exacte n’ont pas de contrainte sur la
10 Résumé
distance inter-élément du réseau d’antennes, donc elles sont applicables aux réseaux
d’antennes avec un espacement entre éléments à demi-longueur d’onde.
En conclusion, dans cette thèse, nous avons proposé deux méthodes de localisation
de sources basées sur le modèle exact, une méthode basée sur le modèle approximé
et deux méthodes de correction pour la localisation de sources en champ proche d’un
système MIMO bistatique.
Contents
Remerciements 3
Résumé 5
List of Figures 15
List of Algorithms 17
List of Acronyms 19
Notations 21
Introduction 23
1 Radar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.2 Passive and Active Radar Systems . . . . . . . . . . . . . . . . . . 24
1.3 Array Based Radar . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2 MIMO Radar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1 Advantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Classification of MIMO Radar Systems . . . . . . . . . . . . . . . 27
3 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4 Organization of this Dissertation . . . . . . . . . . . . . . . . . . . . . . . 32
5 List of Related Publications . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1 Signal Model 35
1.1 Target Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.2 Modeling of Bistatic MIMO System . . . . . . . . . . . . . . . . . . . . . . 36
1.3 Modeling of Near Field Targets . . . . . . . . . . . . . . . . . . . . . . . . 38
1.3.1 Fresnel Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.3.2 Wavefront Approximation in Near Field . . . . . . . . . . . . . . 38
1.3.3 Symmetric Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
1.4 Signal Model in Cartesian Coordinates System . . . . . . . . . . . . . . . 47
11
12 Contents
1.5 Cramér-Rao Lower Bound (CRLB) . . . . . . . . . . . . . . . . . . . . . . 49

1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2 Approximated Model Based Methods 51

2.1 Summary of [1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.2 Extended [2] for MIMO Systems . . . . . . . . . . . . . . . . . . . . . . . 57
2.3 Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3 Correction Methods 65
3.1 Approximation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 LUT Based Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2.1 LUT Creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2.2 Correction using LUT . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3 Correction using Estimated Accurate Relative Propagation Path Length
Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3.1 Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3.2 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.4 Four Location Parameters to Cartesian Coordinates . . . . . . . . . . . . 74
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4 Accurate Model Based Methods 81

4.1 Improved [1] for Accurate Model . . . . . . . . . . . . . . . . . . . . . . . 81
4.2 PARAFAC Based Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Conclusion 91
1 Conclusions and Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . 91
2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Appendices 95
A Biased Location Parameters 95

A.1 Biased Location Parameters Extraction Methods . . . . . . . . . . . . . . 95
A.1.1 Vector Flip . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
A.1.2 Double Difference . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
A.1.3 Overdetermined System of Linear Equations . . . . . . . . . . . . 98
A.1.4 Least Squares Error . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
A.1.5 Vector Component Pairs . . . . . . . . . . . . . . . . . . . . . . . . 99
Contents 13
A.1.6 Residue Vector Approximation . . . . . . . . . . . . . . . . . . . . 100

A.2 Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
B Correction Methods for Single Array 105
Bibliography 111
List of Figures
1 Passive and active radar systems schematic. . . . . . . . . . . . . . . . . 25

2 Types of MIMO radar. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.1 Transmitting ULA parameters . . . . . . . . . . . . . . . . . . . . . . . . . 37

1.2 Fresnel region. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.3 Variation of |ϑr( p, N ) | with respect to θr p when ρr p = F̆r , no = 1, and dr =
λ/2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.4 Variation of |ϑr( p, 1) | with respect to θr p when ρr p = F̆r , no = N, and
dr = λ/2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.5 Variation of |ϑr( p, N ) | with respect to θr p when ρr p = F̆r , no = 1, and dr =
λ/4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.6 Lower bounds of ρr p as a function of N when dr = λ/2. . . . . . . . . . . 43
1.7 Lower bounds of ρr p as a function of N when dr = λ/4. . . . . . . . . . . 43
1.8 Lower bounds of ρr p as a function of N̆ when dr = λ/2 in a symmetric
array. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.9 Lower bounds of ρr p as a function of N̆ when dr = λ/4 in a symmetric
array. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.10 Cone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.11 A bistatic MIMO radar system with linear arrays. . . . . . . . . . . . . . 48
2.1 Variation of unambiguous region of θe p with de . It is same for θr p with

respect to dr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.2 Ambiguous region of ρe p with respect to θe p at de = λ/4. . . . . . . . . . 56
2.3 Ambiguous region of ρr p with respect to θr p at dr = λ/4. . . . . . . . . . 57
2.4 RMSE in range associated with the transmitting array estimated by the
method in [1] and the proposed method versus SNR. The CRLBs are
calculated using (1.41). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.5 RMSE in DOD estimated by the method in [1] and the proposed method
versus SNR. The CRLBs are calculated using (1.41). . . . . . . . . . . . . 63
2.6 RMSE in range associated with the receiving array estimated by the
method in [1] and the proposed method versus SNR. The CRLBs are
calculated using (1.41). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
15
16 List of Figures
2.7 RMSE in DOA estimated by the method in [1] and the proposed method
versus SNR. The CRLBs are calculated using (1.41). . . . . . . . . . . . . 64
3.1 Discretized Fresnel region of an array. . . . . . . . . . . . . . . . . . . . . 67

3.2 A block of discretized 2D near field region. . . . . . . . . . . . . . . . . . 69
3.3 RMSE in the position estimated by the methods described in Chapter 2
and corrected by the proposed method explained in section 3.2 versus
SNR. The CRLBs are calculated using (1.43) and (1.45). . . . . . . . . . . 77
3.4 RMSE in the position estimated by the methods described in Chapter 2
and corrected by the proposed method explained in section 3.3 versus
SNR. The CRLBs are calculated using (1.43) and (1.45). . . . . . . . . . . 78
4.1 Tensorization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 RMSE in the position estimated by the proposed methods described in
section 4.1 and section 4.2 versus SNR. de = dr = λ/2, K = 103 , L = 103 ,
M = 8, N = 9, mo = 1, no = 1, and P = 2. The CRLBs are calculated
using (1.43) and (1.45). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.3 RMSE in the position estimated by the proposed methods described in
section 2.2, section 4.1, and section 4.2 versus SNR. de = dr = λ/4,
K = 103 , L = 103 , M = 7, N = 9, mo = 4, no = 5, and P = 2. The CRLBs
are calculated using (1.43) and (1.45). . . . . . . . . . . . . . . . . . . . . . 90
A.1 Approximation error (ξer p ) profile of the extraction method in section A.1.1;
dr = λ/4, N = 9, and no = 5. . . . . . . . . . . . . . . . . . . . . . . . . . 101
dr = λ/4, N = 9, and no = 5. . . . . . . . . . . . . . . . . . . . . . . . . . 102
dr = λ/4, N = 9, and no = 5. . . . . . . . . . . . . . . . . . . . . . . . . . 102
dr = λ/4, N = 9, and no = 5. . . . . . . . . . . . . . . . . . . . . . . . . . 103
dr = λ/4, N = 9, and no = 5. . . . . . . . . . . . . . . . . . . . . . . . . . 103
dr = λ/4, N = 9, and no = 5. . . . . . . . . . . . . . . . . . . . . . . . . . 104
B.1 ξer p of [2] in noiseless condition; d = λ/4, L = 103 , and N̆ = 4. . . . . . . 106

B.2 ξer p of [3] in noiseless condition; d = λ/4, L = 103 , and N̆ = 4. . . . . . . 106
B.3 ξrp of [2], corrected using section 3.2; d = λ/4, L = 103 , and N̆ = 4. . . . 107
List of Algorithms
2.1 Proposed Approximated Model Based Method. . . . . . . . . . . . . . . . 61

3.1 Look Up Tables Creation Method. . . . . . . . . . . . . . . . . . . . . . . 68
3.2 Look Up Tables Based Correction Method. . . . . . . . . . . . . . . . . . 70
3.3 Estimated Relative Path Length Vector Based Correction Method. . . . . 74
4.1 Submatrix Based Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2 Tensor Decomposition Based Method. . . . . . . . . . . . . . . . . . . . . 88
17
List of Acronyms
CANDECOMP CANonical DECOMPosition

COMFAC COMplex parallel FACtor
CRLB Cramér–Rao Lower Bound
DOA Direction Of Arrival

DOD Direction Of Departure
DTLD Direct TriLinear Decomposition
EM ElectroMagnetic
ESPRIT Estimation of Signal Parameters via Rotational Invariance Techniques
GRAM Generalized Rank Annihilation Method
LS Least Squares
LUT Look Up Table
MATLAB MATrix LABoratory

MIMO Multiple Input Multiple Output
ML Maximum Likelihood
MSE Mean Square Error
MUSIC MUltiple SIgnal Classification
MVDR Minimum Variance Distortionless Response
PARAFAC PARAllel FACtor

PARAFAC2 Parallel factor for cross-product matrices
PARATUCK PARAFAC–TUCKER
PM Propagator Method
RADAR RAdio Detection And Ranging

RARE RAnk REduction
19
20 List of Acronyms
RCS Radar Cross Section

RD-PARAFAC Reduced Dimensional PARAllel FACtor
RMSE Root Mean Square Error
SNR Signal-to-Noise Ratio

SVD Singular Value Decomposition
TALS Trilinear Alternating Least Squares

TLS Total Least Squares
ULA Uniform Linear Array

Notations
a, α, A Scalar
a, α Vector
A Matrix
A 3-way tensor
[·]T Transpose
[·]∗ Conjugate
[·] H Hermitian transpose (conjugate transpose)
[·]† Left pseudo inverse
[·]−1 Inverse of a square matrix
E{·} Expectation
k · kF Frobenius norm
√
j −1
<{ a} Real part of a
={ a} Imaginary part of a
| a| Absolute value of a
arg{ a} Argument of a
∠{ a } Principal value of a
IN N × N identity matrix
JN N × N exchange matrix
0N N × 1 zeros vector
0N N × N zeros matrix
1N N × N ones matrix
1 M× N M × N ones matrix
N1 Set of all natural numbers without zero
Z Set of all integers
R Set of all real numbers
C Set of all complex numbers
card(·) Cardinal number of a set
Hadamard product
⊗ Kronecker product
Khatri-Rao product (column wise Kronecker product)
21
22 Notations
diag{ a} Diagonal matrix with the elements of a as its diagonal elements

diag{ A} Column vector containing the diagonal elements of A
rank{ A} Rank of A
tr{ A} Trace of A
det{ A} Determinant of A
[ a ]i ith component of a
Nc (µ, σ2 ) Complex circular Gaussian distribution with mean µ and variance σ2
Introduction
RAdio Detection And Ranging (RADAR) has been an important research topic
since many decades. Even now, its development is continuing with a steady pace.
To begin, we summarize the transition of radar technology from conventional to Mul-
tiple Input Multiple Output (MIMO) system with its applications, classifications, and
advantages. MIMO system based radars are the recent kinds of radar systems which
have shown better performance as compared to the conventional radar systems. Con-
sequently, the radar systems with MIMO technology are currently a hot topic of re-
search. The development of signal processing techniques linked to the MIMO radar
systems is discussed under the state of the art. And at last, the organization of this
dissertation is provided, followed by the list of related publications.
1 Radar
Radar is a localization system which analyses the received ElectroMagnetic (EM)
signals to detect, locate, and track the objects in its range. It is because, when an EM
wave interacts with an object, the property of the wave changes based on the charac-
teristics of the objects like distance, velocity, shape, and size. By analyzing the echoed
signal, we can extract the information regarding the characteristics of that object. Ac-
cording to the historical records, it all started in 1886, when Heinrich Rudolf Hertz
decided to practically test the EM theory developed by James Clerk Maxwell. He fur-
ther studied the reflection of the radio waves by different dielectric substances. The
frequency of a radio wave may vary from 3 kHz to 300 GHz. However, the preferred
frequency range in radar is 25 MHz to 70 GHz. Inspired by the work of Hertz, many
researchers worked on the radar systems which were less attractive field at that time.
Many successful experiments like radio telegraphy, detection of ships and aircrafts,
radar based altimeter, range finding, etc. were already performed using continuous
and pulsed waveforms. However, the radar development only accelerated in World
War II for military applications. After the war, the research bent towards the civilian
applications. Since then, it has gone through a lot of improvements and it is still an
active field of research. Parallel researches on the electronics and microwave devices
like processors, magnetrons, klystrons, traveling-wave tubes, mixers, filters, etc. have
23
24 Introduction
played a significantly important role in the development of radar [4, 5].
1.1 Applications
The primary objective of inventing the radar was to locate the aircrafts and ships
in World War II. However, after the end of the war the radar is also developed for the
civilian applications. It is now used for navigation. Modern aircrafts are equipped
with altimeter and weather radars to increase safety. Altimeter radar can also be used
to map the ground from air. Weather radar provides a large resource for the weather
forecast. Radar is very useful in navigation when optical visibility is poor. Doppler
radars are used in road safety by detecting speedy vehicles. Radar can also be used
for surveillance to secure the public places like airports, museums, concerts, politically
sensitive conferences, metro trains, etc. Radar is still used by the military to detect,
track, and guide missiles for national security. During the world wars, many land
mines were buried underground which can still be active. Ground penetration radar
can be used to detect those land mines in order to disable them. Not only the land
mines, the ground penetration radar can also be used to explore ancient ruins without
physically exploiting them. Radar also contributes in the researches related to astron-
omy, civil engineering, medicines, etc. [4–6].
1.2 Passive and Active Radar Systems
A radar is distinguished based on its applications, type of waveform used, and its
significant feature. As a result, there are many ways in which a radar system can be
classified. Consequently, here we only discuss the classification concerning the passive
and active radar systems which are shown in Figure 1.
Passive radar system only contains receiver. It simply listens to the EM signals
emitted or reflected by the targets. Since it has no dedicated transmitter, the targets
which do not emit their own signals can not be detected in the absence of any EM
wave broadcasting source. However, this drawback makes it nearly undetectable to
the surveillance radar.
On the other hand, active radar system contains transmitter along with the receiver.
To locate a target, the receiver processes the reflections of the known signal broadcast
by its dedicated transmitter. Since, we know the waveform of the transmitted signal,
we can extract more information about the target from the echo of transmitted known
signals. MIMO systems fall under active radar [7].
1. Radar 25
Target
Target
Receiver
Receiver Transmitter
Broadcasting
Passive Radar Station Active Radar
(a) Passive Radar System (b) Active Radar System
Figure 1 – Passive and active radar systems schematic.
1.3 Array Based Radar

Initial radar systems were using mechanical setup to rotate a single antenna with
directivity to scan the region of interest. Later phased array got attention which can
electrically steer the beam along the region of interest. The main advantage of electrical
steering is the scanning speed. To steer the beam, each antenna of the array is fed with
different weighted signal waveform. The main purpose of these weights is to induce
different phase delays in the transmitted waveforms corresponding to the different
antennas. However, sometimes amplitude is also modified to get some specific beam
shapes. Finding the location of a target is one of the main objectives of the arrays based
radar.
In 1969, J. Capon proposed a high resolution direction finding algorithm using
phased array for seismic application [8]. This method is also known as Minimum
Variance Distortionless Response (MVDR) beamformer, which signifies its working
principle [9]. More than a decade later, inspired by a frequency estimation technique
known as Pisarenko harmonic decomposition [10], spectral MUltiple SIgnal Classi-
fication (MUSIC) technique was developed which is based on the orthogonality be-
tween the signal and noise subspaces [11, 12]. It has higher resolution capability than
MVDR and can provide unbiased estimates of the number of sources, Direction Of Ar-
rivals (DOAs), and polarizations [12]. For a standard linear array, the cost function
in the spectral MUSIC technique can be transformed into a polynomial. If there are P
sources, then the P roots of this polynomial which are inside and closest to the unit
circle provide the DOAs of the P sources [13]. Because of the involvement with the
polynomial roots, this method is often called as root MUSIC. Beside it, a high resolu-
tion low computational complexity subspace based method, referred to as Estimation
of Signal Parameters via Rotational Invariance Techniques (ESPRIT), was proposed in
1986 which uses two identical arrays with displacement invariance [14–16]. The com-
26 Introduction
putation complexity of this method is close to the root MUSIC, however lower than
the spectral MUSIC. Mean Square Error (MSE) of the DOAs estimated by root MUSIC
is lower than that by ESPRIT [17]. In [16], the authors have claimed that ESPRIT has
more robustness with imperfect array than MUSIC method. Subspace based methods
like MUSIC and ESPRIT usually use eigendecomposition which adds high computa-
tional complexity to these techniques. Propagator Method (PM) is a subspace based
method that doesn’t use eigendecomposition, however provides satisfactory perfor-
mance as compared to the MUSIC techniques [18]. RAnk REduction (RARE) is another
subspace based method which shows performance comparable to the other subspace
based methods [19]. There are many other localization methods such as minimum-
norm, linear prediction, etc. for DOA estimation [20–22]. Almost all of the above
mentioned methods focus on the far field region of a standard linear array. However,
they were also developed further to deal with other array geometries and situations.
As a result, there exist many variations of these methods in the literature concerning
sources localization.
Because of the influence of the military applications on the initial development of

the radar, near field sources localization had received less attention as compared to the
far field. However, the technological upgrades in the human civilization have driven
the development of near field sources localization from last few decades along with
the far field sources localization. The wavefront received at the array from a far field
source is often considered as planar whereas the wavefront is approximated as quadric
(quadratic surface) for a near field source. However, the wavefront of a point like
source is always spherical in practical situation [23, 24]. The mathematical expression
of the spherical wavefront has high nonlinearity. Thus, to reduce the nonlinearity of
this expression only first order term of the Taylor expansion of this expression is used
to express the planar wavefront and the first and second orders terms are used for a
near field source that model a quadric wavefront. When an array of antennas is used,
the range and DOA of each near field source are estimated. However, in case of a
far field source only its DOA is usually estimated. The approximation of the spheri-
cal wavefront adds bias to the estimated location parameters (range and DOA). The
biases in the location parameters due to the wavefront approximation decrease with
the increase in range [23]. Consequently in the far field situation, the bais due to the
wavefront approximation is significantly small as compared to the wavelength of the
source carrier signal. However, this bias can’t be ignored in the near field case, if pre-
cision is our prime concern. Regardless of that there exist many near field sources
localization techniques which use quadric wavefront based model [2, 3, 23–28]. And
these techniques are usually analyzed using the received signals generated with the
quadric wavefront based model which can not be considered as a fair testing of their
performance concerning the estimation error in the location parameters.
2. MIMO Radar 27
2 MIMO Radar
The idea of MIMO communication [29–31] was stretched out to MIMO radar [32,
33]. A MIMO system consists of multiple transmitting and receiving antennas. These
antennas may be placed in distributed or colocated manner. Each transmitting an-
tenna emits a waveform orthogonal to the waveform transmitted by the remaining
transmitting antennas. This orthogonality may exist in the signal space, time domain,
or frequency domain [34]. The receiving array receives the superposition of the echoes
of the transmitted signals after their interaction with the targets. The acquired signal at
each receiving antenna is then passed through the matched filters. The orthogonality
between the transmitted signal allows the matched filters to identify the known trans-
mitted signals in the received signal at each receiving antenna [35, 36]. Consequently,
we collect the number of transmitting antennas times the number of receiving anten-
nas data points at a time instance. Because of the advancements in the computation
capabilities, we can apply complicated signal processing techniques on the data col-
lected at the receiver of a MIMO radar [37]. A MIMO radar can also be considered
as an upgrade to the active phased array radars. The major difference between the
two is that a phased array radar transmits weighted versions of the same signal wave-
form whereas in a MIMO radar each transmitting antenna emits linearly independent
waveforms [38].
2.1 Advantages
A MIMO radar is an active radar therefore it inherits the merits and demerits of an
active radar. Multiple paths in a channel may cause fading of the signals. Due to the
multiple transmitting and receiving antennas, this fading phenomenon can be mini-
mized to a great extent [37,39]. The ability to uniquely identify the targets is known as
parameter identifiability. Due to the use of linearly independent transmitted signals,
MIMO radar has waveform diversity which provides high parameter identifiability as
compared to a phased array radar with same number of transmitting and receiving an-
tennas [40]. In other words, a MIMO radar can unambiguously identify more targets
than a phased array radar.
2.2 Classification of MIMO Radar Systems

In a broad sense, a MIMO radar system is categorized as statistical and coherent [34,
41]. In a statistical MIMO radar, the transmitting and receiving antennas are widely
separated from each other. Due to the spare array, statistical MIMO radar can view a
target from many directions and gather more information about that target [42]. Some
targets may have small Radar Cross Section (RCS) for some locations however large
28 Introduction
RCS at the other locations. These spatially diverse targets can not easily hide from a
MIMO radar with widely separated antennas.
In a coherent MIMO radar, the transmitting and receiving antennas are colocated.
Based on the number of colocated antennas arrays and the separation between the
transmitting and receiving arrays, a coherent MIMO radar can be classified as a mono-
static, bistatic, or multistatic. In monostatic MIMO radar, the transmitting and receiv-
ing arrays overlap. Therefore, the DOA and Direction Of Departure (DOD) for a target
are same. If the surface of target facing the radar has small RCS, the monostatic radar
may not detect it. This situation can be overcome by using a bistatic MIMO radar in
which the transmitting array is placed far from the receiving array [37]. Because of the
large separation between the transmitting and receiving arrays, a bistatic MIMO radar
has more degrees of freedom than a monostatic MIMO radar [37]. In a monostatic
MIMO radar, DOA is equal to DOD, therefore the signal model is simple as compared
to a bistatic MIMO radar and as a result, a monostatic MIMO radar requires estimating
only one direction for each target [43–48]. A multistatic MIMO radar contains many
transmitting and receiving arrays. The antennas in these arrays are locally colocated
however the arrays are widely separated. This increases the spatial diversity of the
mutistatic MIMO radar. It can also be considered as a combination of the statistical
and coherent MIMO radars. The colocated antennas in these radar provide high an-
gular resolution and detection sensitivity [32, 49]. The work in this thesis only deals
with a bistatic MIMO radar which has more spatial diversity than a monostatic MIMO
radar and less complexity than a multistatic MIMO radar.
Monostatic
Statistical
(Widely separated antennas)
MIMO radar Bistatic
Coherent
(Colocated antennas)
Multistatic
Figure 2 – Types of MIMO radar.
3 State of the Art

In a passive radar with standard linear array, we only need to estimate DOA which
often requires one parameter estimation technique [8, 9, 11–22]. However in a bistatic
MIMO radar with standard linear arrays, we need to estimate the DODs along with
the DOAs of targets because it is an active radar. To estimate two parameters, we can
3. State of the Art 29
use a two-dimensional estimator or use two one-dimensional estimators. Maximum

Likelihood (ML) is a usual approach to estimate unknown parameters. However, it is
computationally expensive. As a result, in 2006, a 2D unitary ESPRIT based technique
was proposed to jointly estimate the DODs and DOAs that has less computational
complexity than a ML based approach [9, 50]. A two-dimensional estimation method
is generally slower than two one-dimensional estimation methods. Thus, multiple one-
dimensional estimation methods are preferred over one multiple-dimensional estima-
tion method. A drawback of using two one-dimensional estimation methods is that
the DODs and DOAs of the targets are not paired. This lack of pairing can be ob-
served in [51] which uses two 1D ESPRIT to estimate the DODs and DOAs of targets
followed by an additional process to pair them. The pairing process adds additional
computation burden to the whole localization procedure. Furthermore, at low Signal-
to-Noise Ratio (SNR), the paring may not be correct. Keeping these things in mind,
soon after [51], a method was proposed which achieved automatic pairing between
the DODs and DOAs using two 1D ESPRIT with a little modified approach [52]. An-
other study in 2009, achieved automatic pairing using ESPRIT for a signal with colored
noise [53]. However the method was limited to three transmitters. The limitations on
the number of transmitters in [53] were overcame by using Singular Value Decomposi-
tion (SVD) along with ESPRIT in [54]. In 2010, Reduced dimension MUSIC algorithm
was proposed, in which 2D-MUSIC problem was divided into two 1-D MUSIC prob-
lems [55]. Automatic paring of DOD and DOA, can also be achieved using polynomial
rooting technique [56]. In this, polynomial root-MUSIC was compared to 2-D MUSIC
pseudo-spectrum and double 1-D MUSIC pseudo-spectrum and found to be more ef-
ficient than the other two methods. The computation time of root-MUSIC is higher
than ESPRIT. Therefore, a combined solution was proposed [57]. In this method DOD
was estimated by using ESPRIT and their corresponding DOA was estimated by using
root-MUSIC. The performance of the method proposed in [57] was good enough to
extend this method for the polarimetric bistatic MIMO radar [58, 59].
Beside the subspaces based methods, there are factor analysis methods to estimate
DOD, DOA and Doppler shift. Uniqueness and rotation of factors analysis were dis-
cussed in [60–62]. Using factor analysis, Tucker decomposition was proposed to de-
compose a three dimensional tensor into three matrices (2-D tensors) and one 3-D
tensor factors [63]. In 1969, an attempt was made to define the rank of tensors by
taking the rank of matrices as a reference [64]. In 1970, PARAllel FACtor (PARAFAC)
and CANonical DECOMPosition (CANDECOMP) trilinear decompostions were intro-
duced independently in [65] and [66] respectively. In 1972, Harshman introduced
Parallel factor for cross-product matrices (PARAFAC2) [67]. In this, he also discussed
about the algorithm and uniqueness of PARAFAC2 model. J. B. Kruskul extended
the matrix rank to the rank of multi-way array and provided a relation between the
30 Introduction
uniqueness and the rank of 3-way array for PARAFAC [68–70]. In [71], Harshman and
Lundy reviewed on direct and indirect fitting of PARAFAC, PARAFAC2 and PARAFAC–
TUCKER (PARATUCK) models. In 1997, Rasmus Bro published tutorial and applica-
tions of PARAFAC under Chemometrics and Intelligent Laboratory Systems [72]. In
this, he discussed about the uniqueness and rank of multi-way arrays. He also summa-
rized Trilinear Alternating Least Squares (TALS) and three different ways to speed it
up. He used chemometric examples to show his finding. He further elaborated multi-
way analysis for food industry in his PhD thesis [73]. His team also proposed a direct
fitting algorithm for PARAFAC2 [74] and used it in fluorescence spectroscopy [75].
Sensor array processing became an application of PARAFAC [76, 77]. In 1999, COM-
plex parallel FACtor (COMFAC) was introduced which dealt with PARAFAC for tri-
linear complex valued data [78]. University of Copenhagen, has provided free N-way
toolbox for MATrix LABoratory (MATLAB) [79]. This toolbox contains the updated
versions of useful functions, algorithms and documents. To avoid confusions, [80]
proposed standard notations for the tensors and operations on it. Uniqueness, being
an important characteristics of tensor decomposition, was extended from trilinear to
multilinear case [81]. The multi-way analysis has became a part of multilinear (tensor)
algebra [82, 83]. The tensor algebra can be considered as a super set of linear alge-
bra [84, 85].
Many fitting algorithms were proposed for PARAFAC to decompose tensor. How-
ever, among those TALS was found to have better performance [86]. TALS some-
time converges very slowly between initial and final state. Since, TALS convergences
time depends upon the size of tensor, therefore some techniques were proposed using
Tucker method to compress its size [87, 88]. In TALS, the initialization also plays a
significant role in convergence and avoiding local minima. Few initialization method
like Generalized Rank Annihilation Method (GRAM), Direct TriLinear Decomposition
(DTLD) were also helpful to accelerate the convergences [89–92]. Another solution to
speed up the convergence of TALS is line search. Line search was initially used in [65]
with a constant step size (1.25). Later, another iteration dependent step size (cube root
of iteration) was proposed which performed better than constant step size [73]. In 2008,
enhanced line search algorithm for real valued tensors was proposed in which step size
was a factor that minimizes the cost function [93]. Following this research work, the al-
gorithm was extended to complex valued tensors [94]. The above mentioned methods
have greatly improved TALS performance. However, due to the increase in the area of
applications, fast TALS was proposed for multi-way tensors which achieved memory
and time benefits over previous TALS algorithms [95].
The localization using MIMO radar can be considered as a trilinear algebra prob-
lem which was the most common data structure in tensor decomposition research. In
case of bistatic MIMO radar, the performance of PARAFAC was proved much bet-
3. State of the Art 31
ter in terms of computation time and estimation error as compared to 2-D spectral
Capon [96]. In mutistatic and bistatic case, the performance of PARAFAC was found
better than 2-D spectral subspace techniques with considerable resolution [97]. Due to
trilinear decomposition, DOA, DOD and Doppler frequency were estimated jointly for
a bistatic MIMO radar using PARAFAC [98]. In 2014, PARAFAC-MUSIC was proposed
in which the estimated Doppler frequencies were fed again to the matched filters bank
to cancel the effect of Doppler shift and estimated DOA (monostatic) using filtered
signal with the help of MUSIC [99]. PARAFAC-MUSIC showed reduction in estima-
tion error of DOA as compared to PARAFAC and MUSIC. A Reduced Dimensional
PARAllel FACtor (RD-PARAFAC) was proposed for monostaic MIMO radar configu-
ration and compared with Cramér–Rao Lower Bound (CRLB), PARAFAC, ESPRIT and
PM [100]. According to the results in [100], RD-PARAFAC outperformed other meth-
ods and for low SNR, PM showed high estimation error. Tensor-ESPRIT showed less
angle estimation error as compared to PARAFAC [101].
All the above mentioned methods are dedicated to the far field targets localization
using bistatic MIMO radar. Very few methods can be found in the existing literature
which can deal with the localization of near field targets using bistatic MIMO radar
even though near field sources localization techniques using passive array radar are
under constant development [1, 23, 25–28, 102]. In this thesis, we try to fill this gap
related to the targets in the near field region of a bistatic MIMO radar. The near field
region of an array can not be mathematically modeled the same as its far field region
because the far field assumptions in near field region may lead to a significant estima-
tion error. For a point like target, the signal wavefront is spherical which has a non-
linear mathematical expression. Taylor expansion is used on this expression to reduce
its nonlinearity making the localization procedure relatively simple. Based on this ap-
proximated model, [26] presents a subspaces based technique to localize the targets
in the near field region of a pseudo-monostatic MIMO radar. Inspired by their work,
another method was proposed for the near field region of a bistatic MIMO radar [1].
Due to the combination of the near field situation and bistatic MIMO radar, we need
to estimate four localization parameters viz. DOD, DOA, and two ranges belonging to
the transmitting and receiving arrays. The basic strategy in most of the existing meth-
ods for the near field sources localization is to estimate DOAs before the estimation of
their corresponding ranges. This strategy can work fine with the approximated model,
however it won’t work correctly for the accurate model. In this thesis, we present some
novel methods to deal with the two models.
32 Introduction
4 Organization of this Dissertation
This dissertation deals with some localization techniques to estimate the positions
of the targets in the near field region of a bistatic MIMO system. To explain the meth-
ods, we need firstly to provide the signal model. Thus, in chapter 1, we formulate
the signal model of a bistatic MIMO system. We start with a brief explanation of the
modeling of targets in section 1.1, because the echoed signals have space-time varying
random gains due to the interaction with the targets. There exists a huge amount of
work on a bistatic MIMO system, thus, its signal model is easily available in the exist-
ing literature on it. In section 1.3, we explain the model of the near field region of a
bistatic MIMO radar. Near field region, also referred to as Fresnel region, of an array
of antennas is a finite 3D space. This finite region is bounded by the upper and lower
limits of the radial distance from the center of the array. We also give the expressions
to compute these limits in the same section. Approximation of the model is a common
practice while dealing with the Fresnel region. The change in the signal model due
to the Fresnel approximation of the spherical wavefront is elaborated in section 1.3.2.
Another common practice in the near field sources localization methods is the use of
symmetric Uniform Linear Array (ULA) whose mathematical meaning is provided in
section 1.3.3. While locating a near field source using a linear array, its distance from
the array’s reference point and its angle at that reference point with respect to the axis
of the array constitute its location parameters. In a bistatic MIMO system with linear
arrays, we have four such location parameters belonging to the transmitting and re-
ceiving arrays. However, in 3D space we need only three coordinates. Additionally, it
is more practical to visualize and analyze the location of target in terms of Cartesian co-
ordinates. Thus, in section 1.4, we restate the signal model in terms of three Cartesian
coordinates. To analyze the estimation accuracy of an estimator, the Root Mean Square
Error (RMSE) of the estimated parameter is compared against its CRLB which serves
as a benchmark. The expression to calculate the CRLBs of the location parameters of
the targets in the near field region of a bistatic MIMO system can easily be found in the
related work. Section 1.5 is dedicated to the expression used to calculate the CRLB of
the location parameters in the bistatic MIMO case.
Chapter 2 contains two localization methods which use Fresnel approximation based
signal model. Section 2.1 provides a summary of an existing near field target localiza-
tion method using bistatic MIMO system which is used in this work to compare with
our proposed methods. This method has very low computational complexity but the
maximum number of localizable targets solely depends on the number of sensors in the
receiving array. Contrary to it, the proposed method in section 2.2 is computationally
costly however it can locate more number of targets by exploiting the MIMO technol-
ogy. In section 2.3, we compare the RMSE of the four location parameters estimated by
4. Organization of this Dissertation 33
the two methods in the presence of two targets.
As mentioned before, the Fresnel approximation introduces biases in the estimated

location parameters. This bias can be considered as a systematic error which can eas-
ily be removed by calibrating the estimated localization parameters. In chapter 3, we
propose two correction methods to reduce the bias introduced by the Fresnel approxi-
mation. The first method in section 3.2 is based on Look Up Table (LUT) of the biases
in the discretized near field region; the second method in section 3.3 is based on the
estimation of propagation path length vector from the estimated location parameters.
The LUT based correction method is possible because the finite near field region leads
to finite LUT. It is obvious that there can be infinite locations in the finite near field re-
gion. As a result, we discretize the near field region of each array and store the biases
of the discretized spaces in the LUTs. There is no need to update a LUT unless the cor-
responding array parameters change. The arrays parameters, like number of sensors
and interelement distance, are rarely modified in real life applications. The estimated
location parameters of a target related to the transmitting and receiving arrays are cor-
rected using a low complexity interpolation technique. Section 3.2 discusses the LUT
creation and interpolation procedures in detail. An approach to deal with the accurate
model is to estimate directional vectors and then extract the unbiased location param-
eters from them. The second proposed correction method described in section 3.3 uses
this approach. The principle of this correction technique is explained in section 3.3.1
and the correction procedure is given in section 3.3.2. The correction performance of
the proposed methods is analyzed in section 3.5 by correcting the location parameters
estimated by the methods in chapter 2. The analysis is performed by comparing the
standard deviations of the position errors before and after corrections. Cartesian co-
ordinates based signal model allows us to study the estimation performance in terms
of position error. A method to obtain the Cartesian coordinates from the four location
parameters is provided in section 3.4.
In chapter 4, we describe two proposed methods which can directly handle the
accurate model. Both methods use the same strategy in which the directional vec-
tors are estimated firstly and then the unbiased location parameters are extracted from
them. The method in section 4.1 is based on the subspace based technique whereas
the method in section 4.2 uses tensor decomposition to estimate the directional vec-
tors. Section 4.3 shows an example in which we compare the mean squares of position
errors of the proposed methods against CRLB and a method which uses Fresnel ap-
proximation.
Finally, we conclude the work with perspectives and future works.

34 Introduction
5 List of Related Publications

Letter
• Parth Raj Singh, Yide Wang, and Pascal Chargé, “Bistatic MIMO Radar for Near
Field Source Localisation Using PARAFAC”, Electronics Letters 52 (12), 1060-
1061, April 2016. [103]
Articles
• Parth Raj Singh, Yide Wang, and Pascal Chargé, “An Exact Model-Based Method
for Near-Field Sources Localization with Bistatic MIMO System”, Sensors 17 (4),
723, 2017. [104]
• Parth Raj Singh, Yide Wang, and Pascal Chargé, “A Correction Method for the
Near Field Approximated Model Based Localization Techniques”, Elsevier Digi-
tal Signal Processing 67, pp. 76–80, 2017. [105]
• Parth Raj Singh, Yide Wang, and Pascal Chargé, “Performance Enhancement of
Approximated Model Based Near Field Sources Localisation Techniques”, IET
Signal Processing 11(7), pp. 825–829, 2017. [106]
Conferences
• Parth Raj Singh, Yide Wang, and Pascal Chargé, “Near Field Targets Localization
Using Bistatic MIMO System with Symmetric Arrays”, EUSIPCO, pages 2467–
2471, Kos island, Greece, August 2017. [107]
• Parth Raj Singh, Yide Wang, and Pascal Chargé, “Near Field Targets Localization
Using Bistatic MIMO System with Spherical Wavefront Based Model”, EUSIPCO,
pages 2472–2476, Kos island, Greece, August 2017. [108]
1
Signal Model
In this chapter, we discuss the necessary mathematical models useful for under-
standing the context of the following chapters. At first, we introduce the distributions
often used to model a complex target followed by the commonly used model of the
received signal in a bistatic MIMO radar system. Next, we describe the signal model
with Fresnel approximation in the near field region followed by the effect of this ap-
proximation. We also reformulate the signal model in Cartesian coordinates. At last,
we provide the expressions to obtain the CRLBs of the location parameters for the
signal models concerning spherical wavefront, quadric wavefront, and Cartesian coor-
dinates.
1.1 Target Model

In an active radar, the targets are detected by analyzing the reflection of the trans-
mitted signals from targets on the receiving antennas. The reflectivity of a target is
measured in terms of its RCS which depends on the physical dimensions, geometry,
orientation with respect to the incident wave, and constituent material of its surface.
The reflectivity of the material of its exposed surface is a function of frequency [5].
Therefore, many practical radar systems use multiple frequencies to get sufficiently
large RCS in order to increase the probability of detection as compared to the single
frequency case. Regardless of that, in this study, we use a single carrier frequency with
the assumption that RCS of each target is large enough at this particular frequency.
For some geometries of targets, a slight movement in their orientation may generate
scintillation. Based on this, the targets can be categorized into fluctuating or nonfluc-
35
36 Chapter 1. Signal Model
tuating. Nearly spherical objects generate negligible scintillation, thus, they fall under
nonfluctuating targets [109]. The RCS of a fluctuating target is not easily tractable.
As a result, the reflection coefficient of a fluctuating target is modeled as a random
process [110]. Swerling cases are the most recognized models for the fluctuating tar-
gets [34, 109, 110]. Swerling cases I and II are applicable to non spherical targets with
more than one large reflectors as compared to the wavelength of impinging signal like
airplanes. Swerling cases III and IV are used to describe the targets with one large
reflector and some small reflectors like ships. Swerling case zero is for nonfluctuating
targets like meteors. In case of Swerling cases I and III, the power of the reflected signal
from a target is nearly constant for all the pulses in a single scan but greatly changes
from scan to scan. In the case of Swerling cases II and IV, the power fluctuates from
pulse to pulse even in a single scan [109]. These cases are not enough to model the
behavior of every fluctuating target [111]. There exist different distributions to model
different complex targets, among which Rayleigh is mostly opted [4]. Hence, we use
complex targets with Rayleigh distributed reflected signal gain.
1.2 Modeling of Bistatic MIMO System

Let P be the number of narrow-band stationary point sources in the near-field re-
gion of a bistatic MIMO system with ULAs. In the following, M and N represent,
respectively, the number of omnidirectional antennas in the transmitting and receiving
arrays of the bistatic MIMO system.
For such a bistatic MIMO system in the presence of P stationary point targets, the
received matched filtered signal at time t can be written as [34, 41, 112]
y(t) = A s(t) + w(t) (1.1)
with A = Ae Ar where Ae ∈ C M× P and Ar ∈ C N × P contain the directional vectors of

departure and arrival, respectively, s(t) ∼ Nc (0, Rs = diag{[σs21 , σs22 , · · · , σs2P ]}) ∈ CP
is the vector of the reflection coefficients at time t of the complex targets, and w(t) ∼
Nc (0, σw2 IM N ) ∈ C MN is an additive noise vector. Here, σs2p = E{[s] p (t) [s∗ ] p (t)} and
σw2 = E{[w]l 0 (t) [w∗ ]l 0 (t)} with p ∈ {1, 2, · · · , P} and l 0 ∈ {1, 2, · · · , M N }.
Let ae p and ar p respectively be the pth columns of Ae and Ar given by
h iT
ae p = ae (ρe p , θe p ) = ae(1, p) , · · · , ae(mo −1, p) , 1, ae(mo +1, p) , · · · , ae( M, p) (1.2)
and
h iT
ar p = ar (ρr p , θr p ) = ar( p, 1) , · · · , ar( p, no −1) , 1, ar( p, no +1) , · · · , ar( p, N ) (1.3)
1.2. Modeling of Bistatic MIMO System 37
( ρ e p , θ e p , ρr p , θr p )
ρep
ρe p
)
,p
δ e (m
de θe p
1 m mo − 1 mo mo + 1 mo + 2 M
Figure 1.1 – Transmitting ULA parameters
where mo and no are the indexes of the reference elements of the transmitting and re-
ceiving arrays, respectively, ae(m, p) = aem (ρe p , θe p ) = exp(− j 2 π δe(m, p) /λ), and ar( p, n) =
arn (ρr p , θr p ) = exp(− j 2 π δr( p, n) /λ). m ∈ {1, 2, · · · , mo , · · · , M} and n ∈ {1, 2, · · · ,
no , · · · , N } are the indexes of the antennas of the respective arrays. λ is the wavelength
of the carrier. As shown in Figure 1.1, δe(m, p) is the difference between the distance trav-
eled by the transmitted signal from the mth transmitting antenna to the pth target and
the distance traveled by the transmitted signal from the mo th transmitting antenna to
the pth target, which can be expressed for ULA as
r
δe(m, p) = δem (ρe p , θe p ) = ρ2e p + (m − mo )2 d2e − 2 (m − mo ) de ρe p cos θe p − ρe p (1.4)
where ρe p and θe p are respectively the range and DOD of the pth target with respect to
the reference transmitting antenna indexed by mo , and de is the inter-element spacing
in the transmitting ULA. Similarly, δr( p, n) is the difference between the distance traveled
by the reflected signal from the pth target to the nth receiving antenna and the distance
traveled by the reflected signal from the pth target to the no th receiving antenna, which
can be expressed for ULA as
r
δr( p, n) = δrn (ρr p , θr p ) = ρ2r p + (n − no )2 d2r − 2 (n − no ) dr ρr p cos θr p − ρr p (1.5)
where ρr p and θr p are, respectively, the range and DOA of the pth target with respect
to the reference receiving antenna indexed by no and dr is the inter-element spacing in
the receiving ULA [23].
Additionally, let
h iT
δe p = δe (ρe p , θe p ) = δe1 (ρe p , θe p ), δe2 (ρe p , θe p ), · · · , δe M (ρe p , θe p ) ∈ RM (1.6)
and
h iT
δr p = δr (ρr p , θr p ) = δr1 (ρr p , θr p ), δr2 (ρr p , θr p ), · · · , δr N (ρr p , θr p ) ∈ R N . (1.7)
δem (ρe p , θe p ) and δrn (ρr p , θr p ) can be called as accurate relative propagation path lengths,
which make δe (ρe p , θe p ) and δr (ρr p , θr p ), the accurate relative propagation path length
vectors respectively.
1.3 Modeling of Near Field Targets

Based on the target-array distance with respect to the array’s aperture, the target
can be considered to lie in near or far field region of that array. The limit between near
and far fields is defined by Fraunhofer distance in radar signal processing.
1.3.1 Fresnel Region

The lower and upper limit of the Fresnel region of an array can respectively be
calculated as [2, 102]
p
F̆ = 0.62 D3 /λ (1.8)
and
F = 2 D2 /λ (1.9)
where D is the largest dimension of the array shown in Figure 1.2. Bistatic MIMO sys-
tem has two arrays, therefore, we have three near field regions. Two near field regions
correspond to the arrays and one to the whole system. This work mostly focuses on
the shared near field regions of the arrays. For the transmitting and receiving ULAs,
D = ( M − 1) de and D = ( N − 1) dr respectively. Let F̆e and F e be the lower and upper
limits of the Fresnel region of the transmitting array and F̆r and Fr be the lower and
upper limits of the Fresnel region of the receiving array respectively.
1.3.2 Wavefront Approximation in Near Field

Expressions in (1.4) and (1.5) are based on spherical wavefront which is often ap-
proximated to simplify the signal model. Since, (1.4) and (1.5) are similar expressions,
the approximation processes of δe(m, p) and δr( p, n) are same. Hence, we only elaborate
approximation in δr( p, n) in the following.
1.3. Modeling of Near Field Targets 39
2
2D
Far-field region
Far-field region
λ
F=
D
q
F̆ = 0.6
2 D3
λ
Near-field region
Transmitting ULA : D = ( M − 1)de
Receiving ULA : D = ( N − 1) dr
Figure 1.2 – Fresnel region.
Equation (1.5) can be rewritten as

q
δr( p, n) = ρr p 1 + ϑr( p, n) − 1 (1.10)
where
(n − no )2 d2r ( n − n o ) dr
ϑr( p, n) = − 2 cos θr p . (1.11)
ρ2r p ρr p
The square root term of (1.10) is usually approximated by using Maclaurin series
(Taylor series centered at the origin), given by [2, 23, 24]
q 1 1 1
1 + ϑr( p, n) = 1 + ϑr( p, n) − ϑr2( p, n) + ϑr3( p, n) + · · · , ∀ |ϑr( p, n) | < 1 (1.12)
2" 8 16 #
1 (n − no )2 d2r ( n − n o ) dr
= 1+ −2 cos θr p
2 ρ2r p ρr p
"
1 (n − no )2 d2r (n − no )3 d3r
2
− 4 cos θ rp − 4 cos θr p
8 ρ2r p ρ3r p
#
(n − no )4 d4r
+ +··· (1.13)
ρ4r p
( n − n o ) dr ( n − n )2 d2
o r 2
= 1− cos θr p + sin θr p + γr( p, n) (1.14)
ρr p 2 ρ2r p
where γr( p,n) represents the remainder with higher degree terms.
The maximum positive and minimum negative values of n − no can be obtained
when no = 1 and no = N at n = N and n = 1 respectively. We have max{(n −
no )} = N − 1 and min{(n − no )} = 1 − N. Also, ϑr( p, n) is inversely proportional to ρr p .
Therefore, ϑr( p, n) achieves maximum value when ρr p = F̆r . Figure 1.3 and Figure 1.4
show the variation of |ϑr( p, n) | for the two cases at dr = λ/2. It can be observed that
the curves in the two figures are mirror images of each other at θr p = 90◦ . Thus,
exploring n − no > 0 situation is sufficient to understand n − no < 0 situation. For the
completeness, Figure 1.5 shows the variation of |ϑr( p, n) | with θr p for some vital values
of N when dr = λ/4. Form the figures and (1.11), we can say that |ϑr( p, n) | is maximum
at ρr p = F̆r with θr p = 180◦ and max{(n − no )} or θr p = 0◦ and min{(n − no )}.
1.4
|ϑ(p,N) |; ρrp = F̆r and no = 1 N = 20
1
1.2
N = 30
1
N = 40
|ϑr(p, N )| (λ)
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100 120 140 160 180
θrp (degree)
Figure 1.3 – Variation of |ϑr( p, N ) | with respect to θr p when ρr p = F̆r , no = 1, and dr =

λ/2.
The Maclaurin series in (1.12) converges only if |ϑr( p, n) | < 1 or
(n − no )2 d2r ( n − n o ) dr
−1 < − 2 cos θr p < 1 (1.15)
ρ2r p ρr p
If we consider n − no > 0 then with few operations of inequality, we can write
(n − no )2 d2r + ρ2r p (n − no )2 d2r − ρ2r p

> cos θr p > (1.16)
2 ( n − n o ) dr ρr p 2 ( n − n o ) dr ρr p
1.4
N = 20 |ϑ(p,1)|; ρrp = F̆r and no = N
1
1.2
N = 30
1
N = 40
|ϑr(p, 1 )| (λ)
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100 120 140 160 180
θrp (degree)
Figure 1.4 – Variation of |ϑr( p, 1) | with respect to θr p when ρr p = F̆r , no = N, and dr =

λ/2.
1.4
|ϑ(p,N) |; ρrp = F̆r and no = 1
1
1.2
N = 50
N = 60
1
N = 70
|ϑr(p, N )| (λ)
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100 120 140 160 180
θrp (degree)
Figure 1.5 – Variation of |ϑr( p, N ) | with respect to θr p when ρr p = F̆r , no = 1, and dr =

λ/4.
which implies that to keep convergence of (1.12), it is necessary that
(n − no )2 d2r + ρ2r p n o
> max cos θr p (1.17)
2 ( n − n o ) dr ρr p
and
(n − no )2 d2r − ρ2r p n o
< min cos θr p . (1.18)
2 ( n − n o ) dr ρr p

As we know that −1 ≤ cos θr p ≤ 1. Therefore, we can say
ρ2r p + (n − no )2 d2r > 2 (n − no ) dr ρr p (1.19)
and
ρ2r p − (n − no )2 d2r > 2 (n − no ) dr ρr p . (1.20)
Furthermore, ρ2r p + (n − no )2 d2r > ρ2r p − (n − no )2 d2r , thus, inequality (1.20) is a suf-
ficient condition for the convergence of (1.12). This inequality can further be used to
calculate the minimum value of range for which the Fresnel approximation is valid. To
obtain that, we need to rewrite (1.20) as
√ √
ρr p − (1 + 2)(n − no ) dr ρr p − (1 − 2)(n − no ) dr > 0. (1.21)
√
In this inequality, ρr p − (1 − 2)(n − no ) dr > 0 when n − no > 0, therefore ρr p >
√
(1 + 2)(n − no ) dr is a necessary condition for the convergence of (1.12).
In the Fresnel region with Fresnel approximation, ρr p ≥ F̆r > 0 and ρr p > (1 +
√
2)(n − no ) dr for n − no > 0. Figure 1.6 and Figure 1.7 show these bounds as a func-
√
tion of N for dr equal to λ/2 and λ/4 respectively. As compared to F̆r , (1 + 2)(n −
no ) dr is linear and grows slowly with the increase of the number of array elements.
This linear bound can not be ignored in the arrays with small aperture. The similar
conclusion can be made for the transmitting array.
For n − no < 0, the inequality (1.15) can be rearranged as
(n − no )2 d2r + ρ2r p (n − no )2 d2r − ρ2r p

< cos θr p < . (1.22)
2 ( n − n o ) dr ρr p 2 ( n − n o ) dr ρr p
From the approach similar to n − no > 0 situation, we can say that ρ2r p − (n − no )2 d2r >
−2 (n − no ) dr ρr p is a necessary condition for the convergence of (1.12) when n − no <
√
0, which gives ρr p > −(1 + 2)(n − no ) dr . Finally, we can simply say that for the
convergence of (1.12),
√
ρr p > (1 + 2)|n − no | dr (1.23)
60
50
40
ρrp (λ)
30
20
10
F̆r
√
( 2 + 1) (N − 1) dr
0
0 5 10 15 20 25 30 35 40
N
Figure 1.6 – Lower bounds of ρr p as a function of N when dr = λ/2.
45
40
35
30
25
ρrp (λ)
20
15
F̆r
√
10 ( 2 + 1) (N − 1) dr
0
0 10 20 30 40 50 60 70
N
Figure 1.7 – Lower bounds of ρr p as a function of N when dr = λ/4.

is the necessary condition for any possible value of n. The maximum positive and
minimum negative values of n − no can be combined as max{|n − no |} = N − 1. By
using this in (1.23) at ρr p = F̆r , we can get
√
N > λ((1 + 2)/0.62)2 /dr (1.24)
which when substituted with dr = λ/2 and dr = λ/4 gives N > 30.3248 and N >
60.6496. The same results can be graphically validated from Figure 1.3, Figure 1.6, and
Figure 1.7. Here, the value of N decreases with the increase of ρr p and this constraint
√
totally vanishes when ρr p > (1 + 2)( N − 1) dr .
In far field situation, only
the
first degree term of (1.14) is used [41, 97], which gives
δr( p,n) ≈ −(n − no ) dr cos θr p . While in case of near field, where the second degree
term is also used, the approximated relative path length is given by [2, 3, 23, 24]
δer( p, n) = δern (ρr p , θr p ) = −(n − no ) ωr p + (n − no )2 φr p (1.25)

= δr( p, n) − ρr p γr( p, n) (1.26)
where ωr p = dr cos(θr p ) and φr p = d2r sin2 (θr p )/(2 ρr p ). Similarly, δe(m, p) is approxi-
mated as
δee(m, p) = δeem (ρe p , θe p ) = −(m − mo ) ωe p + (m − mo )2 φe p (1.27)
where ωe p = de cos(θe p ) and φe p = d2e sin2 (θe p )/(2 ρe p ).

Let a
ee p and a
er p be the approximated (biased) directional vectors of the pth target
associated with (1.27) and (1.25) respectively, which can be expressed as
h iT
a ee ( ρ e p , θ e p ) = e
ee p = a ae(1, p) , · · · , e
ae(mo −1, p) , 1, e
ae(mo +1, p) , · · · , e
ae( M, p) (1.28)
and h iT
a er (ρr p , θr p ) = e
er p = a ar( p, 1) , · · · , e
ar( p, no −1) , 1, e
ar( p, no +1) , · · · , e
ar( p, N ) (1.29)
where eae(m, p) = exp (− j 2 π δee(m, p) /λ) and e

ar( p, n) = exp (− j 2 π δer( p, n) /λ). Further, we
can rewrite the received matched filtered signal vector, in (1.1), for the approximated
model as
e s(t) + w(t)
ye(t) = A (1.30)
e=A
with A ee er where A
A ee = [ a
ee 1 , a
ee 2 , · · · , a er = [ a
eeP ] and A er1 , a
er2 , · · · , a
erP ].
Similar to (1.6) and (1.7), let
h iT
δee p = δee (ρe p , θe p ) = δee1 (ρe p , θe p ), δee2 (ρe p , θe p ), · · · , δee M (ρe p , θe p ) ∈ R M (1.31)
and
h iT
δer p = δer (ρr p , θr p ) = δer1 (ρr p , θr p ), δer2 (ρr p , θr p ), · · · , δer N (ρr p , θr p ) ∈ R N . (1.32)
Further, we can also write

" #
ωe p
δee p = δee (ρe p , θe p ) = δee (ωe p , φe p ) = M (1.33)
φe p
and " #
ωr p
δer p = δer (ρr p , θr p ) = δer (ωr p , φr p ) = N (1.34)
φr p
where
" #T
−(1 − mo ) −(2 − mo ) · · · 1 0 −1 · · · −( M − mo )
M= ∈ R M ×2
(1 − m o )2 (2 − m o )2 · · · 1 0 1 ··· ( M − m o )2
(1.35)
and
" #T
−(1 − no ) −(2 − no ) · · · 1 0 −1 · · · −( N − no )
N= ∈ R N ×2 . (1.36)
(1 − n o )2 (2 − n o )2 · · · 1 0 1 ··· ( N − n o )2
1.3.3 Symmetric Arrays

It can be observed from the existing work on near field sources localization that
most of the methods use symmetric arrays [2, 3, 23, 24]. Here, symmetry is in the rela-
tive indexes of the antennas with respect to the reference antenna in an array. Mathe-
matically, we can use the following constraints to make the transmitting and receiving
arrays of a bistatic MIMO system symmetric.
1. M ∈ {2 M̆ + 1 : M̆ ∈ N1 }
2. N ∈ {2 N̆ + 1 : N̆ ∈ N1 }
3. mo = ( M + 1)/2
4. no = ( N + 1)/2
Like Figure 1.6 and Figure 1.7, Figure 1.8 and Figure 1.9 are drawn for the sym-
metric arrays. In a symmetric receiving array, no is always in the middle of the arrays,
therefore the maximum positive and minimum negative values coexist in the same
array which are N̆ and − N̆ respectively.
In the figures, it can be observed that the linear bound is not as dominating as in the
case of the asymmetric arrays in Figure 1.6 and Figure 1.7. The reason is that max{|n −
√
no |} = N̆ < N − 1. Like (1.24), we can also write N̆ > λ((1 + 2)/0.62)2 /(8 dr ) for the
18
16
14
12
10
ρrp (λ)
F̆r
2 √
( 2 + 1) N̆ dr
0
1 2 3 4 5 6 7 8 9
N̆
Figure 1.8 – Lower bounds of ρr p as a function of N̆ when dr = λ/2 in a symmetric

array.
4
ρrp (λ)
2 F̆r
√
( 2 + 1) N̆ dr
0
1 2 3 4 5 6 7 8 9
N̆
Figure 1.9 – Lower bounds of ρr p as a function of N̆ when dr = λ/4 in a symmetric

array.
1.4. Signal Model in Cartesian Coordinates System 47
symmetric array which give N̆ > 3.7906 and N̆ > 7.5812 for dr = λ/2 and dr = λ/4
respectively that can also be observed in Figure 1.8 and Figure 1.9. Throughout this
thesis, we consider that the convergence condition of (1.12) is always satisfied.
1.4 Signal Model in Cartesian Coordinates System

Near field target localization using a bistatic MIMO system with ULAs requires the
estimation of four location parameters, viz. DOD, DOA, and the distances from a target
to the transmitting and receiving arrays [1]. However, in 3D space, three parameters
are enough to define a location. In other words, the four location parameters have
some redundancy. As a result, we remodel the directional vectors using three Cartesian
coordinates. However, to do so, we need to assign a position vector to every antenna
in transmitting and receiving arrays. Since, we are dealing with ULAs, the unit vector
along the axis of the array and the position vector of its reference antenna are sufficient
to obtain the position vectors of the remaining antennas of that array. Let eo and ro be
the position vectors of the reference transmitting and receiving antennas respectively
with respect to the origin of the Cartesian coordinates system and dce and dcr are the
unit vectors along the axes of the transmitting and receiving ULAs respectively.
)
z n ge
t (ra gle
)
n
igh
e na
la
nth ect
io
ls a ir
(d
g le
n )
a lf a i s ( u n i t vector
h array ax
sensor
O
base
circular
y
x
Figure 1.10 – Cone
In 3D space, the range and directional angle of a target with respect to a linear array
make a circle related to the base of a cone, shown in Figure 1.10, with the range as its
slant height and the directional angle as its half angle. In the bistatic case, we have two
× νe p
d ce
Transmitter θep
Receiver sin
ρep
Target z

ρe p sin θe p νe p
tp ρe p
eo
θe p
ρr p
θr p
d cr
O
y

ro
νr p
θ rp
ρr p sin

ρr p sin θr p dcr × νr p
dce
x
Figure 1.11 – A bistatic MIMO radar system with linear arrays.
such circles (shown in Figure 1.11). The target is located at the intersection of these two
circles. In the figure, νe p and νr p are unit vectors on the planes of the respective circles.
The parametric equations of the circles can be written as
h i
ψe p ( ϕe ) = ρe p sin θe p cos ( ϕe ) νe p + sin ( ϕe ) dce × νe p + ρe p cos θe p dce + eo
(1.37)
and
h i
ψr p ( ϕr ) = ρr p sin θr p cos ( ϕr ) νr p + sin ( ϕr ) dcr × νr p + ρr p cos θr p dcr + ro
(1.38)
where × denotes the cross-product operation between two vectors; ψe p ( ϕe ) and ψr p ( ϕr )
are the position vectors of a point on the respective circles at ϕe and ϕr , respectively.
The equation parameters ϕe and ϕr independently vary from 0 to 2π rad to completely
sweep the respective circles.
Let t p = [ xt p , yt p , zt p ] T be the position vector of the pth target. Then, the ranges
and directional angles can be expressed in terms of the Cartesian coordinates as ρe p =
1.5. Cramér-Rao Lower Bound (CRLB) 49
kt p − eo k F , ρr p = kt p − ro k F , θe p = arccos[(t p − eo )T dce /ρe p ], and θr p = arccos[(t p −

ro ) T dcr /ρr p ]. Thus, according to (1.2)–(1.5), ae p and ar p can respectively be determined
by t p as

2π
ae(m, p) = aem (t p ) = exp − j
λ
q

t p − eo 2 + (m − mo )2 d2e − 2 (m − mo ) de t p − eo T dc − t p − eo (1.39)
F e F
and

2π
ar( p, n) = arn (t p ) = exp − j
λ
q

t p − ro 2 + (n − no )2 d2r − 2 (n − no ) dr t p − ro T dc − t p − ro . (1.40)
F r F
1.5 Cramér-Rao Lower Bound (CRLB)
CRLB is a lower bound of the variance of an unknown parameter computed by a

minimum variance unbiased estimator which serves as the benchmark to compare the
performance of other estimation techniques used to estimate that unknown parameter
from the signal containing it along with a random variable [113]. The variance corre-
sponding to the CRLB is given by the inverse of the Fischer information. In our case,
we have three or four unknown parameters corresponding to each target. Therefore,
the CRLBs of these parameters are given by the diagonal elements of the inverse of the
Fischer information matrix.
CRLB for the ranges and DOA of multiple near field sources has already been de-
rived in [24] from [114]. We can directly use their closed form expression by making
some minor modifications to adapt the four location parameters of each target due to
the use of a bistatic MIMO system. The submatrix of the inverse of the Fischer informa-
tion matrix corresponding to the desired parameters (i.e. the four location parameters
of P targets) can be expressed as [24]
T −1
σw2 H ⊥ H −1
CRLB(η) = < Å Π A Å 14 ⊗ Rs A R A Rs (1.41)
2L
where CRLB(η) ∈ R4P×4P , η = [ρe1 , · · · , ρeP , θe1 , · · · , θeP , ρr1 , · · · , ρrP , θr1 , · · · , θrP ] T
is the vector of the desired location parameters of P targets, L is the number of data
samples, R = A Rs A H + σw2 I M N , Π⊥ †
A = I M N − A A , 14 is the 4 × 4 matrix whose all
components are equal to 1, and
Å = [∂( ae1 ⊗ ar1 )/∂ρe1 , · · · , ∂( aeP ⊗ arP )/∂ρeP , ∂( ae1 ⊗ ar1 )/∂θe1 , · · · ,
∂( aeP ⊗ arP )/∂θeP , ∂( ae1 ⊗ ar1 )/∂ρr1 , · · · , ∂( aeP ⊗ arP )/∂ρrP ,
∂( ae1 ⊗ ar1 )/∂θr1 , · · · , ∂( aeP ⊗ arP )/∂θrP ] . (1.42)
The main diagonal of (1.41) contains the lower bounds of all the location parameters
in terms of variance for all the targets based on the accurate model. Let ε(η p ) be the
standard deviation of η p ∈ {ρe p , θe p , ρr p , θr p }. The CRLB for the approximated model
can simply be obtained by replacing the accurate directional vectors with the approx-
imated directional vectors in (1.41) and (1.42). However, for the signal model with
Cartesian coordinates, (1.41) changes to
T −1
σw2 H ⊥ H −1
CRLBc (ηc ) = < Åc Π A Åc 13 ⊗ Rs A R A Rs (1.43)
2L
where CRLBc (ηc ) ∈ R3P×3P , ηc = [ xt1 , · · · , xtP , yt1 , · · · , ytP , zt1 , · · · , ztP ] T is the vec-
tor of the desired Cartesian coordinates of P targets, 13 is the 3 × 3 matrix of ones,
and
Åc = [∂( ae1 ⊗ ar1 )/∂xt1 , · · · , ∂( aeP ⊗ arP )/∂xtP , ∂( ae1 ⊗ ar1 )/∂yt1 , · · · ,
∂( aeP ⊗ arP )/∂ytP , ∂( ae1 ⊗ ar1 )/∂zt1 , · · · , ∂( aeP ⊗ arP )/∂ztP ] . (1.44)
The CRLBs of the Cartesian coordinates of the pth target can be combined as
q
ε(t p ) = [diag{CRLBc (ηc )}] p + [diag{CRLBc (ηc )}] P+ p + [diag{CRLBc (ηc )}]2P+ p
(1.45)
1.6 Conclusion
The signal models formulated in this chapter are going to serve as the starting point
of the discussion of the methods in the coming chapters. In the next chapter, both the
methods use Fresnel approximation based model. Consequently, the approximated
model is used throughout the next chapter to discuss and analyze the methods.
2
Approximated Model Based Methods
In this chapter, we present an existing approximated model based method and we

propose a new approximated model based method to locate the targets in the near field
region of a bistatic MIMO system with symmetric ULAs. Since the both methods use
the approximated model, we use the approximated signal model given in the previous
chapter to describe the methods and CRLB expression with the approximated model
parameters for the performance analysis of these methods.
2.1 Summary of [1]
In this section, we provide a summary of the method recently proposed in [1]. The
received signal vector, ye(t) in (1.30), can also be expressed as ye(t ) = [ý1T (t), ý2T (t), · · · ,
T ( t ) , · · · , ý T ( t )] T with
ým o M
er D
ým (t) = A e m s(t) + ẃm (t) (2.1)
nh io
e
where Dm = diag ae(m, 1) , ae(m, 2) , · · · , ae(m, P)
e e e and ẃm (t) is the corresponding noise
sub-vector. The method in [1] is a subspace based method which needs symmetric
transmitting and receiving arrays with de ≤ λ/4 and dr ≤ λ/4.
Along with the above constraints, the method in [1] uses M = 5 transmitting an-
tennas to explain their method. Consequently, we also use the same number of trans-
mitting antennas in its summary. The method begins with the construction of four
H ( t )} such that m 6 = m . The four cross
cross covariance matrices, Ŕm = E{ým (t) ým o o
51
52 Chapter 2. Approximated Model Based Methods
covariance matrices are as follows:
Ŕ1 = E{ý1 (t) ý3H (t)} = A

er D erH ,
e 1 Rs A (2.2)
Ŕ2 = E{ý2 (t) ý3H (t)} = A
er D erH ,
e 2 Rs A (2.3)
Ŕ4 = E{ý4 (t) ý3H (t)} = A
er D erH ,
e 4 Rs A (2.4)
Ŕ5 = E{ý5 (t) ý3H (t)} = A
er D erH
e 5 Rs A (2.5)
where E{ẃm (t) ẃ3H (t)} = 0 N for m 6= 3 with 0 N beign a N × N square matrix contain-
ing only zeros. From the first two and last two pairs of the cross covariance matrices,
we can write
−1
er D
Ŕ1 = A e 2∗
e1 D erH A
A er erH Ŕ2
A (2.6)
−1
er D
Ŕ5 = A e 4∗
e5 D erH A
A er erH Ŕ4
A (2.7)
which can further be written as

−1
Ŕ1 Ŕ2+ er D
=A e 2∗
e1 D erH A
A er erH Ŕ2 Ŕ+
A (2.8)
2
−1
Ŕ5 Ŕ4+ = A e5 D
er D e 4∗ erH A
A er erH Ŕ4 Ŕ+
A (2.9)
4
−1
where Ŕ2 Ŕ2+
= Ŕ4 Ŕ4+ e e
= Ar Ar ArH e erH and [·]+ denotes the pseudo inverse
A
operation using singular value decomposition which is explained below.
As mentioned in [1], the inverse of Ŕm should be calculated by using the singular
values and vectors corresponding to the P largest singular values to improve the ro-
bustness in a noisy environment. Because for P < N, Ŕm is rank deficient therefore
Ŕ− 1 H
m does not exist. Hence, Ŕ m is used to represent its inverse. Let Ŕ m = Um Σm Vm
+
where Um ∈ C N × N , Σm ∈ R N × N , and Vm ∈ C N × N are the matrices containing the

left singular vectors, singular values, and right singular vectors respectively. Then,
−1 H N × P , Σ́ ∈ R P× P , and V́ ∈ C N × P are the matrices
m = V́m Σ́m Úm where Úm ∈ C
Ŕ+ m m
that contain the left singular vectors associated with the P largest singular values, the
corresponding P singular values and right singular vectors respectively. In theory, the
remaining ( N − P) singular values will be zero.
er , the following ex-
By right multiplying both sides of the above equations with A
pressions can be obtained
Ŕ1 Ŕ2+ A
er = A
er D e 2∗
e1 D (2.10)
and
Ŕ5 Ŕ4+ A
er = A
er D e 4∗ .
e5 D (2.11)
e1 D
From (2.10) and (2.11), it is clear that D e ∗ and D
e5 D
e ∗ correspond to the P largest
2 4
2.1. Summary of [1] 53
eigenvalues of Ŕ1 Ŕ2+ and Ŕ5 Ŕ4+ respectively. Let u p and u p be the eigenvectors of
Ŕ1 Ŕ2+ and Ŕ5 Ŕ4+ associated to their P largest eigenvalues respectively. The subspace
spanned by [u1 , u2 , · · · , u P ] and [u1 , u2 , · · · , u P ] is also spanned by the columns of
Ar , thus, we have u p = q p aer p and u p = q p a er p , where q p and q p are the constants which
appear during the eigenvalue decomposition.
From the P largest eigenvalues of Ŕ1 Ŕ2+ and Ŕ5 Ŕ4+ , we can have
e 2∗ = diag {[exp(− j 2 π (ωe + 3 φe )/λ), exp(− j 2 π (ωe + 3 φe )/λ),

e1 D
D 1 1 2 2
· · · , exp(− j 2 π (ωeP + 3 φeP )/λ)]} . (2.12)
and
e 4∗ = diag {[exp(− j 2 π (−ωe + 3 φe )/λ), exp(− j 2 π (−ωe + 3 φe )/λ),

e5 D
D 1 1 2 2
· · · , exp(− j 2 π (−ωeP + 3 φeP )/λ)]} . (2.13)
Further from (2.12) and (2.13), the following can be obtained

∗
D e 2∗
e1 D e 4∗ = diag {[exp( j 4 π ωe /λ), exp( j 4 π ωe /λ), · · · , exp( j 4 π ωe /λ)]} .
e5 D
D 1 2 P
(2.14)
+ +
However, before using (2.14), the eigenvalues of Ŕ1 Ŕ2 and Ŕ5 Ŕ4 should be paired by
comparing the inner product of the corresponding eigenvectors from the fact that the
inner product of two aligned vectors is greater than that of two nonaligned vectors [1].
The DOD of the pth target can be calculated from the estimated value of ωe p obtained
from the arguments of the diagonal elements in (2.14) as

θ̂e p = arccos ω̂e p /de . (2.15)
Similar to (2.14), the following can also be written

∗
D e 2∗ D
e1 D e 4∗
e5 D = diag {[exp( j 12 π φe1 /λ), exp( j 12 π φe2 /λ), · · · , exp( j 12 π φeP /λ)]}
(2.16)
From the arguments of the diagonal elements in (2.16), φe p can be estimated and then
the range to the transmitter can be calculated by using the following relation

ρ̂e p = d2e sin 2
θ̂e p /(2 φ̂e p ). (2.17)
To estimate the angle of arrival, let ŭ p = u p JN u∗p = q p q∗p ( a er∗p ), which

er p JN a
gives
h iT
ŭ p = q p q∗p exp(−4 π j N̆ ωr p /λ), · · · , 1, exp(4 π j ωr p /λ), · · · , exp(4 π j N̆ ωr p /λ)
(2.18)
where  
0 0 ··· 0 1
 
0 0 · · · 1 0
 
. .. . .. .. 
JN =  .. . . . . . ∈ RN×N (2.19)
 
 
0 1 · · · 0 0
1 0 ··· 0 0
is the exchange matrix.
From (2.18), ωr p can be estimated as
N −1
λ
ω̂r p =
4 π ( N − 1) ∑ ∠ [ŭ p ]i /[ŭ p ]i+1 . (2.20)
i =1
Then, the angle of arrival is given by

θ̂r p = arccos ω̂r p /dr (2.21)
Let uep = u∗p JN u∗p = q∗p q∗p ( a

er∗p JN a
er∗p ), which gives
h iT
uep = q∗p q∗p exp(4 π j N̆ 2 φr p /λ), · · · , 1, exp(4 π j φr p /λ), · · · , exp(4 π j N̆ 2 φr p /λ)
(2.22)
From (2.22), φr p can be estimated as
N̆
λ
φ̂r p =
8 π N̆
∑ arg [uep ] N̆ −i+1 /[uep ] N̆ +1 + arg [uep ] N̆ +i+1 /[uep ] N̆ +1 /i2 . (2.23)
i =1
The angles should be unwrapped before using (2.20) and (2.23) to remove the phase
ambiguities which appear because the argument of a complex number can only be
retrieved in its principal value. The process of unwrapping angles is given in [115].
In the unwrapping process, the discontinued principal values of the components of uep
are added or subtracted by a multiple of 2π to obtain a smooth transition along the
phases. As we know that the argument of the reference component should be zero.
The remaining unwrapped phases can be shifted according to it to get the arguments
of the components of uep which can be used in (2.23) to compute φ̂r p . Finally, the range
to the receiving array can be computed as

ρ̂r p = d2r sin2 θ̂r p /(2 φ̂r p ). (2.24)
2.1. Summary of [1] 55
In this method, Ŕ1 Ŕ2+ and Ŕ5 Ŕ4+ are N × N complex valued matrices, therefore,
the maximum number of localizable targets is N. Beside, there are some other con-
straints in this method like quarter wavelength inter element spacing. These extra
constraints appear due to the fact that the argument of a complex number can only
be retrieved in its principal form belonging to (−π, π ]. This impacts the parameter
retrieval from (2.14), (2.16), (2.18), and (2.22). For (2.14) and (2.16), the constraints are
−π < 4 π ωe p /λ ≤ π and −π < 12 π φe p /λ ≤ π which are directly visible in the
expression. On manipulating these constraints further, we can get

λ λ
arccos < θe p ≤ arccos − (2.25)
4 de 4 de
and
ρe p ≥ 6 d2e sin 2
θe p /λ. (2.26)
All the elements of ŭ p contain the information about a single target. To retrieve
the necessary information from ŭ p , [1] proposed to use difference between the phases
of its consecutive elements which should be unambiguous. Thus, we can say −π <
4 π ωr p /λ ≤ π is a constraint that can be rewritten as

λ λ
arccos < θr p ≤ arccos − . (2.27)
4 dr 4 dr
Like the elements of ŭ p , the elements of uep also belong to a single target. To avoid
ambiguity in this case, the arguments of the elements adjacent to the reference element
should be between −π and π, i.e. −π < 4 π φr p /λ ≤ π. Accordingly, we can write

ρr p ≥ 2 d2r sin2 θr p /λ. (2.28)
The constraints on the DOD and DOA, given in (2.25) and (2.27) respectively, are
similar. It can be found out from the expressions that the unambiguous sweeps of
θe p and θr p will decrease if the inter element spacings increase beyond λ/4 to reduce
the mutual coupling between array antennas. Figure 2.1 shows the variation of the
unambiguous region of θe p with respect to de varying from λ/4 to λ/2. Below λ/4,
0 ≤ θe p ≤ π and 0 ≤ θr p ≤ π because the lower and upper limits of DOD and DOA
are physically controlled by the array geometry. Consequently, the method in [1] can
only work for the inter element spacings less than or equal to λ/4.
At de = dr = λ/4, (2.26) and (2.28) are the functions of only θe p and θr p respec-
tively. Figure 2.2 and Figure 2.3 respectively show the unambiguous region of ρe p and
ρr p with respect to θe p and θr p at quarter wavelength. From the figures, we can say
that the ambiguities in the ranges are maximum at θe p = θr p = 90◦ because of their
squared sine function. However, their magnitudes differ. By making comparison with
Figure 2.1 – Variation of unambiguous region of θe p with de . It is same for θr p with

respect to dr .
0.7
(M = 5)
0.6
Ambiguous region
0.5 F̆e
(M = 4)
0.4
ρep (λ)
0.3
(M = 3)
0.2
0.1 (M = 2)
0
0 20 40 60 80 100 120 140 160 180
θep (degree)
Figure 2.2 – Ambiguous region of ρe p with respect to θe p at de = λ/4.

2.2. Extended [2] for MIMO Systems 57
0.25
(N = 3)
0.2 Ambiguous region

F̆r
0.15
ρrp (λ)
0.1
(N = 2)
0.05
0
0 20 40 60 80 100 120 140 160 180
θrp (degree)
Figure 2.3 – Ambiguous region of ρr p with respect to θr p at dr = λ/4.
the lower limit of the Fresnel region, we can conclude that the ambiguity in the range
corresponding to transmitting array doesn’t exist for M ≥ 4 and the ambiguity in the
range corresponding to receiving array doesn’t exist for N ≥ 3. As mentioned in the
beginning of this section, the method in [1] uses symmetric arrays, therefore N will
always be greater than or equal to 3. And also M = 5. Hence, the ambiguity due to the
range will never occur while using this method.
2.2 Extended [2] for MIMO Systems
In this section, we propose a new method which can be considered as an extension

of the method in [2] to deal with MIMO systems. As the previously presented method
[1], the proposed one also uses symmetric arrays. [2] uses two subarrays, however, the
division of the array into two subarrays is unnecessary. Thus, in the proposed method,
we use the whole array and its permuted version.
The symmetry in the transmitting ULA allows us to write the following relation
 
exp(− j 2 π (− M̆ ωe p + M̆2 φe p )/λ)
 .. 
 
 . 
 
JM a
ee p 
= 1  = de (θe ) a
ee p (2.29)
 p
 .. 
 . 
 
exp(− j 2 π ( M̆ ωe p + M̆2 φe p )/λ)
where J M is the M × M exchange matrix, is the Hadamard product, and

 
exp( j 4 π M̆ ωe p /λ)
 .. 
 
 . 
 

de (θe p ) =  1 . (2.30)

 . 
 .. 
 
exp(− j 4 π M̆ ωe p /λ)
As mentioned in [2], de ≤ λ/4 is a necessary condition to avoid the phase ambiguity

in the elements of de (θe p ).
By using the above relation, we can write
e = ( De
J̆ A e
1N×P ) A (2.31)
where J̆ = J M ⊗ IN , IN is the N × N identity matrix, 1 N × P is the N × P dimensional

matrix with all its components equal to one, and De = [de (θe1 ), de (θe2 ), · · · , de (θeP )].
The covariance matrix of ye(t) is given by
e = E{ye(t) yeH (t)} ∈ C M N × M N

R
=A e H + σw2 I M N .
e Rs A (2.32)
e can be written as
The eigendecomposition of R
e = Us Λs UsH + σw2 Un UnH

R (2.33)
where the diagonal elements of Λs ∈ RP× P are the P largest eigenvalues and the
columns of Us ∈ C M N × P are their corresponding eigenvectors spanning the signal
subspace. The columns of Un ∈ C M N ×( M N − P) span the noise subspace. From (2.32)
and (2.33), we can write
Ae T = Us (2.34)
with T ∈ CP× P being an invertible square matrix.

2.2. Extended [2] for MIMO Systems 59
For an arbitrary angle θ, we can write
Fe (θ ) = J̆ Us − [de (θ ) ⊗ 1 N × P ] Us (2.35)
e T − [de (θ ) ⊗ 1 N × P ] A
= J̆ A eT
= ( De 1 N × P ) A e T − [de (θ ) ⊗ 1 N × P ] A
eT
= ([de (θe1 ) − de (θ ), de (θe2 ) − de (θ ), · · · , de (θeP ) − de (θ )] 1 N × P ) Us (2.36)
from (2.31) and (2.34) (see [2] and [116]). At θ = θe p , all the components of the pth
column of Fe (θ ) become zero and Fe (θ ) becomes rank deficient. Thus, we can use the
following spectrum function to estimate DODs [2, 116]
1
Se ( θ ) = . (2.37)
det{ FeH (θ ) Fe (θ )}
The estimated DODs (say θ̂e p ) of the P targets correspond to the P highest peaks of
Se (θ ) when θ is varied from 0◦ to 180◦ .
Using the orthogonality between the signal and noise subspaces, we can write [59]
h iH h i
ee ( ρ e p , θ e p ) ⊗ a
a er (ρr p , θr p ) Un UnH ee ( ρ e p , θ e p ) ⊗ a
a er (ρr p , θr p ) = 0. (2.38)
Since, h i
ee ( ρ e p , θ e p ) ⊗ a
a er (ρr p , θr p ) = aee ( ρ e p , θ e p ) ⊗ I N aer (ρr p , θr p ). (2.39)
Hence, we can also say [59]

h iH h i
det ee ( ρ e p , θ e p ) ⊗ I N
a Un UnH ee ( ρ e p , θ e p ) ⊗ I N
a =0 (2.40)
i.e. the subspace spanned by the directional vector of departure is orthogonal to the
noise subspace, therefore, the estimated range corresponding to θ̂e p can be given by
1
ρ̂e p = arg max (2.41)
F̆e ≤ρ≤ F e
det{ G H (ρ) G (ρ)}
where G (ρ) = UnH [ a ee (ρ, θ̂e p ) ⊗ IN ] ∈ C( MN − P)× N . In (2.41), when MN − P ≥ N and

ρ 6= ρe p , G (ρ) has full rank and det{ G H (ρ) G (ρ)} is not zero. At ρ = ρe p , det{ G H (ρ)
G (ρ)} tends towards zero due to the orthogonality between the noise subspace and
the subspace spanned by the directional vector of departure corresponding to the pth
target.
Like the transmitting array, the receiving array is also a symmetric ULA. Therefore,
we can write
e = (1 M× P Dr ) A
Ĭ A e (2.42)
where Ĭ = I M ⊗ JN and Dr = [dr (θr1 ), dr (θr2 ), · · · , dr (θrP )] with

 
exp( j 4 π N̆ dr cos(θr p )/λ)
 .. 
 
 . 
 

dr ( θr p ) =  1 . (2.43)

 . 
 .. 
 
exp(− j 4 π N̆ dr cos(θr p )/λ)
dr ≤ λ/4 is a necessary condition to avoid the phase ambiguity in the elements of

Dr [2].
Like before, the DOAs can be estimated by using the following spectrum function
1
Sr ( θ ) = . (2.44)
det{ FrH (θ ) Fr (θ )}
where
Fr (θ ) = Ĭ Us − [1 M× P ⊗ dr (θ )] Us (2.45)
= (1 M × P [dr (θr1 ) − dr (θ ), · · · , dr (θrP ) − dr (θ )]) Us (2.46)
can be obtained from (2.34) and (2.42). The estimated DOAs (say θ̂r p ) of the P targets
correspond to the P highest peaks of Sr (θ ) when θ is varied from 0◦ to 180◦ . However,
the DOAs are not paired.
Finally, to get the estimation of the receiver side range of the pth target, we use 2D
MUSIC as
1
(ρ̂r p , θ̂r p ) = arg max H
(2.47)
F̆ ≤ρ≤ F
a
er (ρ, θ ) Q a
er (ρ, θ )
r r
θ ∈{θ̂r1 ,··· ,θ̂r P }
where Q = [ a ee (ρ̂e p , θ̂e p ) ⊗ IN ] H Un UnH [ a

ee (ρ̂e p , θ̂e p ) ⊗ IN ]. Here, θ is chosen from the
set of unpaired DOAs which reduces the computational cost. 2D MUSIC automatically
pairs the transmitting and receiving sides location parameters for each target. Because
when Q belongs to the pth target, the pair of range and DOA which minimizes the 2D
MUSIC cost function will correspond to that pth target.
Initially, all the P DODs and DOAs are estimated in one 1D search each. And then,
the ranges belonging to transmitting and receiving arrays of each target are estimated
separately by handling one target at a time. Therefore, all the location parameters are
automatically paired. Algorithm 2.1 provides the complete procedure of this proposed
method.
2.3. Numerical Example 61
Algorithm 2.1 Proposed Approximated Model Based Method.

1: Estimate the covariance matrix Re of the matched filtered received signals ye(t) using
L samples.
2: Separate the eigenvectors corresponding to the signal and noise subspaces using
e
the P largest eigenvalues of R.
3: Estimate θe p and θr p by finding P largest peaks of Se (θ ) and Sr (θ ) respectively.
4: For each estimated θe p , estimate ρe p by using (2.41).
5: For each estimated ρe p and θe p , calculate Q and use it in (2.47) to estimate ρr p and
pair estimated θr p .
The maximum number localizable targets in this method is N ( M − 1), which is

bounded due to the denominator of the cost function in (2.41).
2.3 Numerical Example

Consider a bistatic MIMO system with M = 9 ( M̆ = 4) transmitting and N = 11
( N̆ = 5) receiving antennas. The reference antennas indexes are mo = 5 and no = 6 to
get symmetric ULAs. The arrays elements are uniformly spaced by de = dr = λ/4. λ
is used as the unit of length throughout the simulation. To analyze the performance of
the methods described in section 2.1 and section 2.2, we show the variations in RMSEs
of the location parameters with respect to SNR when the methods process L = 103
signal data temporal samples. The RMSE of each location parameter is calculated by
using K = 103 Monte Carlo trials in the following formula
v
u K
u 1 2
e ηp = t
K ∑ η̂ p (k ) − η p (2.48)
k =1
where η p ∈ {ρe p , θe p , ρr p , θr p } is the true value of the location parameter of the pth
target and η̂ p (k) ∈ {ρ̂e p (k), θ̂e p (k), ρ̂r p (k ), θ̂r p (k)} is the estimated value of the corre-
sponding location parameter in the kth trial.
Almost all of the existing approximated model based methods [1–3, 23, 24] simu-
late the received signals by using Fresnel approximation to analyze the performance
of their methods. The methods in section 2.1 and section 2.2 are based on the approx-
imated model, thus, we also use the model corresponding to (1.30) as the input signal
to show the performance of our method.
In Figure 2.4, Figure 2.5, Figure 2.6, and Figure 2.7, we compare the RMSE in the
location parameters (range associated with the transmitting array, DOD, range associ-
ated with the receiving array, and DOA respectively) of two targets estimated by the
proposed method and the method in [1] in terms of SNR. The SNR is varied from 0 dB
to 30 dB with an interval of 5 dB. The location parameters (ρe p , θe p , ρr p , θr p ) of the two
targets are (1.5λ, 50◦ , 2.5λ, 140◦ ) and (3.5λ, 110.5◦ , 5λ, 60◦ ). The CRLBs of the loca-
tion parameters are calculated from (1.41) by replacing the accurate model parameters
with the approximated model parameters. In the figures, ε(η p ) denotes the CRLB of
the η p parameter belonging to the pth target in terms of standard deviation.
0
10
−1
10
RMSE (λ)
−2
10
ǫ(ρe1 ); Section 2.1

−3 ǫ(ρe2 ); Section 2.1
10
ε(ρe1 ); Section 1.5
ε(ρe2 ); Section 1.5
−4
10
0 5 10 15 20 25 30
SNR (dB)
Figure 2.4 – RMSE in range associated with the transmitting array estimated by the
method in [1] and the proposed method versus SNR. The CRLBs are calculated using
(1.41).
From the above figures, we can observe that the proposed method has better per-
formance in terms of RMSE than that of the method in [1]. The primary reason of the
poor performance of the method in [1] is that it does not exploit all the available infor-
mation. Additionally, the RMSEs of the ranges estimated by the proposed method are
very close to their respective CRLBs.
2.4 Conclusion
In this chapter, we have proposed an extension of the method [2] to localize near
field targets using a bistatic MIMO system consisting of symmetric transmitting and
receiving ULAs. Compared to the existing method in [1], the proposed method has
better performance in terms of RMSE because we exploit all the available information.
2.4. Conclusion 63
1
10
0
10
−1
10
RMSE (degree)
−2
10
−3
10 ǫ(θe1 ); Section 2.1
ǫ(θe2 ); Section 2.1
ǫ(θe1 ); Section 2.2
−4
10 ǫ(θe2 ); Section 2.2
ε(θe1 ); Section 1.5
ε(θe2 ); Section 1.5
−5
10
0 5 10 15 20 25 30
SNR (dB)
Figure 2.5 – RMSE in DOD estimated by the method in [1] and the proposed method
versus SNR. The CRLBs are calculated using (1.41).
2
10
ǫ(ρr1 ); Section 2.1
ǫ(ρr2 ); Section 2.1
1 ǫ(ρr1 ); Section 2.2
10 ǫ(ρr2 ); Section 2.2
ε(ρr1 ); Section 1.5
ε(ρr2 ); Section 1.5
0
10
RMSE (λ)
−1
10
−2
10
−3
10
0 5 10 15 20 25 30
SNR (dB)
Figure 2.6 – RMSE in range associated with the receiving array estimated by the
method in [1] and the proposed method versus SNR. The CRLBs are calculated using
(1.41).
0
10
−1
10
RMSE (degree)
−2
10
−3
10
ǫ(θr1 ); Section 2.1
−4
10 ǫ(θr1 ); Section 2.2
ε(θr1 ); Section 1.5
ε(θr2 ); Section 1.5
−5
10
0 5 10 15 20 25 30
SNR (dB)
Figure 2.7 – RMSE in DOA estimated by the method in [1] and the proposed method
versus SNR. The CRLBs are calculated using (1.41).
Along with it, the proposed method automatically pairs all the four location parame-
ters. [1] uses submatrices of the covariance matrix, therefore, the maximum number of
localizable targets is limited by the number of receiving antennas, i.e. N. However, the
maximum number of localizable targets in the proposed method is given by N ( M − 1).
Both the methods discussed here use Fresnel approximation which introduces biases
in the estimated location parameters. This bias is a function of location which remains
constant for an array if its parameters are not modified. Consequently, this bias can be
considered as an offset and can be removed by calibration. Hence, in the next chapter,
we propose two methods to reduce this bias from the location parameters.
3
Correction Methods
In the previous chapter, we have elaborated two methods which are based on the
approximated signal model which signifies quadric wavefront. However, in practice
the wavefront is spherical. Thus, the location parameters estimated by every approx-
imated model based method is biased. This bias is of systematic type and therefore
its effect can be reduced. In this chapter, we propose two approaches to reduce the
effects of Fresnel approximation on the location parameters. The first method is based
on LUTs whereas the second method uses Least Squares (LS) estimation.
3.1 Approximation Error
From (1.26), we can observe that the approximated model is biased. This bias is like
a location parameters dependent systematic error added to the accurate model. Thus,
when a biased approximated model based method has to deal with the unbiased accu-
rate model, this bias shows up in the estimated location parameters and decreases their
precisions. Testing an approximated model based method with the accurate model
based input signals is perfectly fair because the wavefront of a point source is spher-
ical in reality. In the considered bistatic MIMO system, there are two ULAs which go
through Fresnel approximation independently. However, the effect of this approxima-
tion is the same for both the arrays. As a result, explaining one explains the other.
65
66 Chapter 3. Correction Methods
3.2 LUT Based Correction
In the near field region, LUT based correction is possible in the first place because it
is finite. A LUT based correction method is one of the most primitive but an effective
way to minimize the approximation bias from the estimated location parameters. A
key point to note is that this bias in an estimated location parameter is the result of the
propagation of the systematic error, introduced by the Fresnel approximation, through
the localization process of an approximated model based method. The final result of
an error propagation depends on the path which the error takes. Since, two different
localization techniques use two different algorithms, therefore, the path taken by the
systematic error to reach location parameters will be different. Consequently, different
approximation biases will be observed in the same location parameters for different lo-
calization techniques which can be seen in Appendix A where six different approaches
to obtain biased location parameters are compared. However, if we are looking for one
method to reduce the effect of the Fresnel approximation, LS based LUT inspired by
section A.1.3 can be a right choice which is described in the following.
The whole LUT based correction method can be divided into two major steps. First
step is the creation of LUT, which only depends on the system’s physical parameters.
The second step is the correction of the location parameters using the LUT created in
the first step.
3.2.1 LUT Creation
For a bistatic MIMO system, we need to construct two 2D LUTs to map the true
location parameters with the corresponding biased location parameters obtained due
to the use of the approximated model. A LUT should be finite even if a bounded space
has infinite locations. Thus, to construct a LUT, we need to calculate the biased location
parameters at some discrete locations in the near field region of an array by varying the
range from lower to upper Fresnel limit and the direction angle from 0 to π by some
small intervals. Figure 3.1 shows a discretized near field region of a linear array when
the range and directional angle are sampled by constant intervals. The intersection
points, which can be seen in Figure 3.1, of the curves corresponding to the range and
directional angle inside the Fresnel region are the locations whose biased parameters
are recorded in the LUT.
For the transmitting array, let ρeq and θ eq be the range and DOD at the qth intersec-
tion point (or discrete location) respectively such that F̆e ≤ ρeq ≤ F e and 0 ≤ θ eq ≤ π.
The next sensitive step is to calculate the biased location parameters from δe (ρeq , θ eq ) by
minimizing its distance with δee (ρ , θ e ), like an approximated model based technique.
eq q
3.2. LUT Based Correction 67
Intersection point
Fresnel region
Array
π 0
Figure 3.1 – Discretized Fresnel region of an array.
In other words,
2

eρ eq , e
θ eq = arg min δee (ρe , θe ) − δe (ρeq , θ eq ) (3.1)
(ρe , θe ) F
where eρeq and e

θ eq are respectively the biased range and DOD at the qth discrete location.
The cost function in (3.2) can easily be solved using an iterative method like Newton’s
minimization. However, an iterative approach is often time consuming. Thus, to ob-
tain a closed form expression based solution, we modify the cost function in (3.2) by
using (1.33) that gives
" # 2
ω
e ) = arg min M
e eq , φ e
(ω eq − δe ( ρ e , θ e q .
) (3.2)
(ωe , φe )
q
φe F
The closed form solution of (3.2) in the LS sense can be given by

" #
e eq
ω
e = M † δe (ρeq , θ eq ). (3.3)
φ eq
Finally,
e e
θ eq = arccos ω eq /de (3.4)
and
e 2 2 e e
ρeq = de sin θ eq / 2 φeq . (3.5)
Similar to the transmitting array, a separated LUT can be created for the receiving
array by replacing the transmitting array side parameters with the receiving array side
parameters in (3.3), (3.4), and (3.5). Algorithm 3.1 summarizes the LUTs creation steps.
Algorithm 3.1 Look Up Tables Creation Method.

1: Discretize the near field region of the transmitting and receiving arrays.
2: Construct the accurate relative propagation path length vector for every discretized
location using (1.6) and (1.7).
3: Calculate the biased location parameters for every discretized location using (3.3),
(3.4), and (3.5) for the transmitting array and similar expressions for the receiving
array.
3.2.2 Correction using LUT
The aim of the correction method is to get back the true location parameters from
the biased location parameters estimated by an approximated model based localization
technique. Therefore, when using the LUT in the correction method, eρeq and e
θ eq will be
the inputs, and ρeq and θ eq will be the outputs of the LUT.
Figure 3.2 shows one block of the grid shown in Figure 3.1 inside which we have the
biased location parameters, corresponding to the transmitting array, of the pth target,
when estimated by an approximated model based localization method. In Figure 3.2,
(ρe p , θe p ) is the pair of true transmitting array side location parameters of the pth tar-
ρ e1 , e
get and (ρee p , θee p ) is its corresponding biased location parameters. (e ρ e2 , e
θ e1 ), (e θ e2 ) ,
ρ ,e
(ee3 θ ), and (e
e3 ρ ,e
e4 θ ) are four biased location parameters from the LUT, obtained
e4
using the method described in section 3.2.1, that are adjacent to (ρee p , θee p ). (ρe1 , θ e1 ),
(ρe2 , θ e2 ), (ρe3 , θ e3 ), and (ρe4 , θ e4 ) are the corresponding true location parameters reg-
istered in the LUT. deq represents the distance between (ρee , θee ) and (ep p ρ ,e
θ e ) where
eq q
q ∈ {1, 2, 3, 4}. It is calculated as

r
2 2
deq = eρeq − ρee p + e θ eq − θee p . (3.6)
From Figure 3.2, de1 , de2 , de3 , and de4 can be viewed as four minimum distances be-
tween the biased location parameters in LUT and (ρee p , θee p ). For these four nearest
ρ − ρ ,e
points, [e eq θ eq − θ eq ] T , ∀ q ∈ {1, 2, 3, 4}, can be considered as an approximation
eq
error vector. The main idea is to interpolate the approximation error vector corre-
sponding to the pth target by using the known approximation error vectors at the four
nearest neighbors. One of the low complexity interpolation methods suitable for such
a situation is the natural neighbor interpolation with inverse distance weighting [117].
In our case, the mathematical expression of the natural neighbor interpolation can be
3.2. LUT Based Correction 69
ρ e2 , e
(e θ e2 )
( ρ e2 , θ e2 )
ρ e3 , e
(e θ e3 )
( ρ e3 , θ e3 )
de2
de3
(ρee p , θee p )
(ρe p , θe p )
de1 de4
ρ e4 , e
(e θ e4 )
( ρ e1 , θ e1 ) ρ e1 , e
(e θ e1 )
( ρ e4 , θ e4 )
Figure 3.2 – A block of discretized 2D near field region.
written as " # " #

ρ̂e p − ρee p 4 ρ eq − e
ρ eq
= ∑ dq (3.7)
θ̂e p − θee p q =1 θ eq − e
θ eq
where ρ̂e p and θ̂e p are respectively the corrected range from the transmitting array and
DOD and !
4
−1
dq = deq ∑ 1/deq0 (3.8)
q =1
0
represents the inverse distance weighting. When deq = 0, for one value of q (q ∈
{1, 2, 3, 4}); ρ̂e p = ρeq and θ̂e p = θ eq . It is impossible that two or more distances are
simultaneously zero.
From (3.7), the corrected location parameters can be calculated as

 
" # " # " #  d1 
ρ̂e p ρee p ρ −e ρ ρ −e ρ ρ −e ρ ρ e4 − e
ρ e4  d 2 
= + e1 ee1 e2 ee2 e3 ee3  .
  (3.9)
θ̂e p θee p θ e1 − θ e1 θ e2 − θ e2 θ e3 − θ e3 θ e4 − e
θ e4  d 3 
d4
Similar to the transmitting array side location parameters correction using LUT, the
receiving array side location parameters can be corrected by using the same approach.
Algorithm 3.2 provides the steps to make corrections using LUTs.
Algorithm 3.2 Look Up Tables Based Correction Method.

1: Estimate the DOD, DOA, and ranges of targets by an approximated model based
method.
2: Search in the LUT belonging to the transmitting array for the four nearest neigh-
bors of (ρee p , θee p ) on the basis of the distances calculated using (3.6).
3: Calculate the corresponding weights from the four distances using (3.8).
4: Use (3.9) to obtain the corrected location parameters (ρ̂e p , θ̂e p ).
5: Repeat the above steps from 2 to 4 for the receiving array side range and DOA with
similar mathematical expressions.
3.3 Correction using Estimated Accurate Relative Propa-

gation Path Length Vector
3.3.1 Principle
Consider a single near field target with location parameters (ρeo , θeo , ρro , θro ) and
reflection coefficient s(t). Corresponding to it, the received signal can be written as
y(t) = [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] s(t) + w(t). (3.10)
The noise subspace associated to y(t), in (3.10), can be given by
1
Πn = I MN − [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] H . (3.11)
MN
An accurate model based subspace technique finds the DOA (θr ), DOD (θe ), and
ranges (ρe and ρr ) of the target such that
[ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] H Πn [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] = 0. (3.12)
The solution of (3.12) is given by ρe = ρeo , θe = θeo , ρr = ρro , and θr = θro , which
3.3. Correction using Estimated Accurate Relative Propagation Path Length Vector 71
means

H 1
[ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] I MN − [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] (3.13)
MN

H
[ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]
= [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] H [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]

1
− [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] H [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]
MN
[ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] H [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]
1
= MN − M2 N 2 = 0. (3.14)
MN
An approximated model based subspace technique finds the DOA, DOD, and ranges
by minimizing the following cost function:
(ρee , θee , ρer , θer ) = arg min [ a er (ρr , θr )] H Πn [ a

ee ( ρ e , θ e ) ⊗ a ee ( ρ e , θ e ) ⊗ a
er (ρr , θr )]
( ρ e , θ e , ρr , θr )
(3.15)

= arg min [a er (ρr , θr )] H [ a
ee ( ρ e , θ e ) ⊗ a ee ( ρ e , θ e ) ⊗ a
er (ρr , θr )]
( ρ e , θ e , ρr , θr )
1
− [a er (ρr , θr )] H [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]
ee ( ρ e , θ e ) ⊗ a
MN
[ ae (ρeo , θeo ) ⊗ ar (ρro , θro )] H [ a ee ( ρ e , θ e ) ⊗ a
er (ρr , θr )]

1
= arg min MN − ee ( ρ e , θ e ) ⊗ a
[ a er (ρr , θr )] H
( ρ e , θ e , ρr , θr ) MN
2

[ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]
2

= arg max [ a er (ρr , θr )] H [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]
ee ( ρ e , θ e ) ⊗ a
( ρ e , θ e , ρr , θr )
(3.16)
= arg min k[ a er (ρr , θr )] − [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]k2F .
ee ( ρ e , θ e ) ⊗ a
( ρ e , θ e , ρr , θr )
(3.17)
Ideally, the minimum of the cost function in (3.17) should be equal to zero. How-
ever, from (1.2), (1.3), (1.28), and (1.29), we can say that ae (ρeo , θeo ) 6= a ee (ρeo , θeo ) and
ar ( ρro , θro ) 6 = a
er (ρro , θro ), or, ae (ρeo , θeo ) ⊗ ar (ρro , θro ) and a
ee ( ρ e o , θ e o ) ⊗ a
er (ρro , θro ) are
not collinear. As a result, the problem becomes to find (ρee , θee , ρer , θer ) such that the
correlation between ae (ρeo , θeo ) ⊗ ar (ρro , θro ) and a

ee ( ρ e , θ e ) ⊗ a
er (ρr , θr ) maximizes:

(ρee , θee , ρer , θer ) = arg max [ a er (ρr , θr )] H [ ae (ρeo , θeo ) ⊗ ar (ρro , θro )]
ee ( ρ e , θ e ) ⊗ a
( ρ e , θ e , ρr , θr )
(3.18)

M N

= arg max ∑ ∑ exp j 2 π δeem (ρe , θe ) + δern (ρr , θr )
( ρ e , θ e , ρr , θr ) m =1 n =1

− δem (ρeo , θeo ) − δrn (ρro , θro ) /λ . (3.19)

The cost function in (3.19) is a nonlinear maximization problem, therefore, its solution
in a closed form expression is difficult to achieve. However, two suboptimal solutions
can be found as
M −1 2 2
e
(ρee , θee ) = arg min ∑ δem (ρe , θe ) − δem (ρeo , θeo ) = arg min δee (ρe , θe ) − δe (ρeo , θeo )
F
(ρe , θe ) m =0 (ρe , θe )
(3.20)
and
N −1 2 2
e
(ρer , θer ) = arg min ∑ δrn (ρr , θr ) − δrn (ρro , θro ) = arg min δer (ρr , θr ) − δr (ρro , θro ) .
F
( ρr , θr ) n =0 ( ρr , θr )
(3.21)
It is obvious from (3.20) that
2 2
e e e
min δe (ρe , θe ) − δe (ρeo , θeo ) = δe (ρee , θe ) − δe (ρeo , θeo )
(ρe , θe ) F F
2

≤ δee (ρeo , θeo ) − δe (ρeo , θeo ) (3.22)
F
which means that δee (ρee , θee ) is a better estimate of the true δe (ρeo , θeo ) than δee (ρeo , θeo )
and that (ρee , θee ) is a biased estimation of (ρeo , θeo ). δee (ρee , θee ) being an optimal estimate
of δe (ρeo , θeo ) in LS sense, a better estimate of (ρeo , θeo ) can be expected from it. The
same can be said for the receiving array side range and DOA from (3.21).
3.3.2 Procedure
Let (ρee p , θee p , ρer p , θer p ) be the location parameters of the pth target estimated by an
approximated model based method. δem (ρe p , θe p ) and δrn (ρr p , θr p ) can respectively be
estimated by substituting these values in (1.27) and (1.25) as
δ̂e(m, p) = δeem (ρee p , θee p ) (3.23)

3.3. Correction using Estimated Accurate Relative Propagation Path Length Vector 73
and
δ̂r( p, n) = δern (ρer p , θer p ). (3.24)
On replacing δem (ρe p , θe p ) by δ̂e(m, p) in the rearranged form of (1.4), we can get

2 (m − mo ) de ρe p cos θe p + 2 δ̂e(m, p) ρe p = (m − mo )2 d2e − δ̂e2(m, p) (3.25)
For all values of m, we can construct an overdetermined system of M linear equations

with ρe p cos(θe p ) and ρe p as the unknowns. Since, δ̂e(m, p) appears on the either sides
of (3.25), Total Least Squares (TLS) is a better approach to solve this system of linear
equations than the conventional LS according to [118]. In order to obtain TLS solutions,
let ve p ∈ R3 be the right-singular-vector associated to the smallest singular value of the
following matrix constructed from the coefficients of (3.25):
 
2 (1 − m o ) d e 2 δ̂e(1, p) (1 − mo )2 d2e − δ̂e2(1, p)
 
 2 (2 − mo ) de 2 δ̂e(2, p) (2 − mo )2 d2e − δ̂e2 
 (2, p) 
 .. .. .. 
 . . . 
 
 2 2 2
. (3.26)
 2 (m − mo ) de 2 δ̂e(m, p) (m − mo ) de − δ̂e 
 (m, p) 
 .. .. .. 
 . . . 
 
2 2 2
2 ( M − mo ) de 2 δ̂e( M, p) ( M − mo ) de − δ̂e( M, p)
From ve p , the estimates of the transmitting array side range and DOD of the pth target
can be calculated as ρ̂e p = −[ve p ]2 /[ve p ]3 and θ̂e p = arccos([ve p ]1 /[ve p ]2 ) respectively.
Similar to (3.25), replacing δrn (ρr p , θr p ) by δ̂r( p, n) in the rearranged form of (1.5) leads
to
2 (n − no ) dr ρr p cos θr p + 2 δ̂r( p, n) ρr p = (n − no )2 d2r − δ̂r2( p, n) (3.27)
which can be used to construct an overdetermined system of N linear equations in

ρr p cos(θr p ) and ρr p for all the values of n. Like above, let vr p ∈ R3 be the right-singular-
vector associated with the smallest singular value of the following matrix constructed
from the coefficients of (3.27):
 
2 (1 − n o ) dr 2 δ̂r( p, 1) (1 − no )2 d2r − δ̂r2( p, 1)
 
 2 (2 − no ) dr 2 δ̂r( p, 2) (2 − no )2 d2r − δ̂r2 
 ( p, 2) 
 .. .. .. 
 . . . 
 
 . (3.28)
 2 (n − no ) dr 2 δ̂r( p, n) (n − no )2 d2r − δ̂r2 
 ( p, n) 
 .. .. .. 
 . . . 
 
2 2 2
2 ( N − no ) dr 2 δ̂r( p, N ) ( N − no ) dr − δ̂r( p, N )
The receiving array side range and DOA of the pth target can respectively be calculated
by ρ̂r p = −[vr p ]2 /[vr p ]3 and θ̂r p = arccos([vr p ]1 /[vr p ]2 ).
The complete correction procedure is summarized in Algorithm 3.3.
Algorithm 3.3 Estimated Relative Path Length Vector Based Correction Method.
1: Estimate the DOA, DOD, and ranges of targets by an approximated model based
technique.
2: Calculate δ̂e(m, p) and δ̂r( p, n) for the pth target using (3.23) and (3.24).
3: Construct the matrices (3.26) and (3.28) from δ̂e(m, p) and δ̂r( p, n) respectively.
4: Using TLS method, get ρ̂e p , θ̂e p , ρ̂r p , and θ̂r p of the pth target.
5: Repeat the above steps from 2 to 4 for each target.
3.4 Four Location Parameters to Cartesian Coordinates

As mention in section 1.4 that three independent coordinates are sufficient to define
a position in a 3D space and analyzing the four location parameters individually is also
not very practical. Hence, we convert the four redundant location parameters to three
independent Cartesian coordinates so that we can easily calculate the localization error
in terms of the Euclidean distance between the true and estimated positions of a target.
The estimated DOA, DOD, and ranges of a target can be used in (1.37) and (1.38) to
construct the parametric equations of the circles corresponding to that target as
h i
ψe p ( ϕe ) = ρ̂e p sin θ̂e p cos ( ϕe ) νe p + sin ( ϕe ) dce × νe p + ρ̂e p cos θ̂e p dce + eo
(3.29)
and
h i
ψr p ( ϕr ) = ρ̂r p sin θ̂r p cos ( ϕr ) νr p + sin ( ϕr ) dcr × νr p + ρ̂r p cos θ̂r p dcr + ro.
(3.30)
As shown in Figure 1.11, the required coordinates are at the intersection of these
circles. However, due to the estimation error and noise, the circles may not intersect;
thus, we propose to minimize the following cost function:
2

( ϕ̂e , ϕ̂r ) = arg min ψe p ( ϕe ) − ψr p ( ϕr ) (3.31)
( ϕe , ϕe ) F
2

where ψe p ( ϕe ) − ψr p ( ϕr ) can be interpreted as the square of the Euclidean distance
F
between the two points on the two circles in 3D space.
A coarse solution of (3.31) can be calculated by exhaustive grid search, and then it
can be finely tuned by Newton’s minimization method. Finally, the position vector of
3.5. Numerical Example 75
the pth target can be computed as t̂ p = ψe p ( ϕ̂e ), t̂ p = ψr p ( ϕ̂r ), or the average of these
two position vectors.

For the simulations in this section, we generate the signal data from the accurate
signal model explained in section 1.4. Then, we use the methods described in chapter 2
to obtain the four location parameters from this accurate signal model data. Because of
the Fresnel approximation, the location parameters estimated from the methods have
some biases. We try to reduce this bias from each location parameter by using the
proposed correction methods in this chapter. Consequently, we have four corrected
location parameters with redundant information for each target. Thus, we compute
the corresponding Cartesian coordinates from the method provided in section 3.4. Fi-
nally, we express the localization error as the Euclidean distances between the true and
estimated positions of a target to analyze the correction performance of the proposed
methods.
Assume a bistatic MIMO system, as shown in Figure 1.11, with M = 7 ( M̆ = 3)
transmitting and N = 9 ( N̆ = 4) receiving antennas. The methods discussed in chap-
ter 2 require symmetric ULAs with inter element spacings de = dr = λ/4. Therefore,
mo = 4 and no = 5. Like before, here too, λ is considered as the unit of length. To
use the signal model in section 1.4, we need the following additional parameters of the
bistatic MIMO system:
1. eo = [0.6λ, λ, 1.4λ] T ,
2. ro = [0.5λ, −0.4λ, −0.9λ] T ,

p
3. dce = [cos(80◦ ), cos(70◦ ), 1 − cos2 (80◦ ) − cos2 (70◦ )] T , and
p
4. dcr = [cos(40◦ ), cos(110◦ ), 1 − cos2 (40◦ ) − cos2 (110◦ )] T .
The values of these parameters are arbitrarily chosen such that there exists a near field
region of significant size which is shared by both ULAs.
Unlike in section 2.3, the performance of the proposed correction methods elabo-
rated in this chapter is analyzed as the root of mean squares of the Euclidean distances
between the true and K estimated positions of a target, with respect to SNR, calculated
as v
u
u 1 K 2
e t p = t ∑ t̂ p (k ) − t p F (3.32)
K k =1
where K is the total number of Monte Carlo trials at each SNR and t̂ p (k) is the estimated
position of the pth target at the kth trial.
In Figure 3.3, we compare e(t p ) of two near field targets whose location parame-
ters are estimated by the methods described in chapter 2 and corrected using the LUT
based correction method given in section 3.2. The position vectors of the two targets
are [4λ, 3λ, 2λ] T and [−2.9λ, 1.7λ, 1.5λ] T . Figure 3.4 shows the similar plots when the
correction is performed using the proposed method elaborated in section 3.3.
From Figure 2.4 to Figure 2.7, we can conclude that the proposed method in sec-
tion 2.2 provides better precision than the existing method summarized in section 2.1.
The same can be observed in Figure 3.3 and Figure 3.4 when the input signals are mod-
eled without Fresnel approximation. The presence of systematic bias in the estimated
position due to the Fresnel approximation can be seen in both the figures as the floor
effect at high SNR in the curves corresponding to the method in section 2.2 before
correction. At low SNR, this systematic bias is not distinguishable in the estimated po-
sition because the noise contributes more to the localization error than the systematic
bias. The proposed correction methods only reduce the effect of the systematic bias
added during the Fresnel approximation. The estimation errors in the positions esti-
mated by the method in section 2.1 for the two chosen targets are so large that the floor
effect is not visible in the figures.
It can be observed in both the figures that both of the proposed correction meth-
ods have high performance when the estimates are good, because the performance of
each correction method also depends upon the performance of the used approximated
model based method. Also, both the proposed correction methods have almost sim-
ilar correction performance. However, the computation complexity of the method in
section 3.3 is much lower than the LUT based correction method in section 3.2.
Additionally, the two proposed correction methods are applicable to a single receiv-
ing array. Therefore, they can also be used for the near field sources localization tech-
niques based on the approximated model such as [2] and [3]. Appendix B shows the
further analysis of the correction methods on such approximated model based meth-
ods.
3.6 Conclusion
The proposed correction methods can significantly reduce the systemic biases ap-
peared in the estimated location parameters due to the Fresnel approximation at high
SNR when the effect of the bias dominates the additive noise. The correction perfor-
mance of both methods improves with the decrease in the estimation error. In the sim-
ulation, a numerical example with symmetric ULAs having quarter wavelength inter
element spacings is considered. However, the proposed correction methods have no
such requirements. In other words, the proposed correction methods can be applied
to an approximated model based method which uses non-symmetric arrays with inter
3.6. Conclusion 77
1
10
0
10
RMSE (λ)
−1
10
−2
10
−3
10
0 5 10 15 20 25 30
SNR (dB)
Figure 3.3 – RMSE in the position estimated by the methods described in Chapter 2 and
corrected by the proposed method explained in section 3.2 versus SNR. The CRLBs are
calculated using (1.43) and (1.45).
The legends of the figure are as follows:
e(t1 ); Section 2.1
e(t2 ); Section 2.1
e(t1 ); Section 2.1 & Section 3.2
e(t1 ); Section 2.2
e(t2 ); Section 2.2
ε(t1 ); Section 1.5
1
10
0
10
RMSE (λ)
−1
10
−2
10
−3
10
0 5 10 15 20 25 30
SNR (dB)
Figure 3.4 – RMSE in the position estimated by the methods described in Chapter 2 and
corrected by the proposed method explained in section 3.3 versus SNR. The CRLBs are
calculated using (1.43) and (1.45).
The legends of the figure are as follows:
e(t1 ); Section 2.1
e(t2 ); Section 2.1
e(t1 ); Section 2.2
e(t2 ); Section 2.2
3.6. Conclusion 79
element spacings less than or equal to λ/2. Fresnel approximation is the cause of the
biases in the location parameter, therefore, by avoiding it, we can directly obtain unbi-
ased location parameters. In the next chapter, we propose two methods which do not
use Fresnel approximation.
4
Accurate Model Based Methods
Beside reducing the bias due to the Fresnel approximation, we can directly exploit
the accurate model. In this chapter, we discuss two proposed methods which can di-
rectly deal with the accurate signal model. The key strategy used by both the proposed
methods is to estimate δe p and δr p first, then use the accurate model on them to obtain
the estimates of ρe p , θe p , ρr p , and θr p . The first proposed method in this chapter is an im-
provement of [1] to deal with the accurate signal model. The second proposed method
in this chapter uses a tensor decomposition technique known as PARAFAC to estimate
the directional vectors.
4.1 Improved [1] for Accurate Model

As in [1], the received signal vector is subdivided into M subvectors belonging to
each transmitting antenna as y(t) = [y1T (t), y2T (t), · · · , y TM (t)] T . The subvector corre-
sponding to the mth transmitter can be written as
ym (t) = y̆m (t) + wm (t) (4.1)
where wm (t) is the corresponding noise subvector and y̆m (t) = Ar Dm s(t) where
nh io
Dm = diag ae(m, 1) , ae(m, 2) , · · · , ae(m, P) (4.2)
is the diagonal matrix with the elements of the mth row of Ae as its diagonal compo-
nents.
81
82 Chapter 4. Accurate Model Based Methods
The eigendecomposition of the covariance matrix R = E{y(t) y H (t)} ∈ C MN × MN

can be written as
R = U Λ UH (4.3)
where the columns of U are the eigenvectors and the diagonal elements of Λ are the
corresponding eigenvalues. If the eigenvalues are sorted in descending order, then the
qth eigenvalue can be expressed as

µ̆ + σ2 if q ∈ {1, 2, · · · , P}
q w
µq = (4.4)
 σ2 if q ∈ { P + 1, P + 2, · · · , MN }
w
where µ̆q is the qth eigenvalue among the P nonzero eigenvalues of A Rs A H arranged
in descending order with Rs = E{s(t) s H (t)}. The following subspace based approach
can be used to remove the additive noise, in which the noiseless covariance matrix is
reconstructed from U and Λ of R as [119]
R̆ = Us diag{[µ1 − σw2 , µ2 − σw2 , · · · , µ P − σw2 ]}UsH (4.5)
where Us contains the columns of U corresponding to the P largest eigenvalues of

R. When a finite number of samples is used, then µ P+1 6= µ P+2 6= · · · 6= µ MN 6=
σ2 . To overcome it, we estimate the noise variance by σ̂w2 = (µ P+1 + µ P+2 + · · · +
µ MN )/( MN − P).
Theoretically, R̆ = A Rs A H can be viewed as a block matrix with submatrices
H N×N
R̆(m, m0 ) = E{y̆m (t) y̆m 0 ( t )} ∈ C
∗ H
= A r Dm R s Dm 0 Ar (4.6)
where m0 ∈ {1, 2, · · · , mo , · · · , M }. Because of the N × N dimension of R̆(m, m0 ) , the

maximum number of targets this method can localize is N. When m = mo , we have
∗ H
R̆(mo , m0 ) = Ar Rs Dm 0 Ar (4.7)
because Dmo is an identity matrix. Using (4.7), (4.6) can be rewritten as
R̆(m, m0 ) = Ar Dm Ar† R̆(mo , m0 ) . (4.8)
On rearranging, we can get
R̆(m, m0 ) R̆+
(m o, m
0) = Ar Dm Ar† (4.9)
where [•]+ denotes the inverse operation using singular value decomposition which is
4.1. Improved [1] for Accurate Model 83
explained below.
As mentioned in [1], the inverse of R̆(mo , m0 ) should be calculated by using the sin-
gular values and vectors corresponding to the P largest singular values to improve
the robustness in a noisy environment. It is because, when P < N, R̆(mo , m0 ) should
theoretically be a non-invertible square matrix. Thus, we use the following method to
obtain its inverse and use R̆+ (mo , m0 )
to represent its inverse. Let R̆(mo , m0 ) = Um0 Sm0 VmH0
where Um0 ∈ C N × N , Sm0 ∈ R N × N , and Vm0 ∈ C N × N are the matrices containing the
left singular vectors, singular values, and right singular vectors respectively. Then,
−1 H N × P , S̆ 0 ∈ R P× P , and V̆ 0 ∈ C N × P are the ma-
R̆+
(mo , m0 )
= V̆m0 S̆m 0 Ŭm0 where Ŭm0 ∈ C m m
trices that contain the left singular vectors associated with the P largest singular values,
the corresponding P singular values and right singular vectors respectively. In prac-
tice, the remaining ( N − P) singular values will be close to zero. Therefore, R̆(mo , m0 )
will not be rank deficient and the conventional way of inverting a square matrix may
introduce numerical accuracy. Thus, this step also improves the numerical accuracy of
the calculation of pseudo-inverse.
The submatrices with the same index m have the same Dm , thus we can add them
together as
1 M
M m∑
Rm = R̆(m, m0 ) R̆+
(mo , m0 )
= Ar Dm Ar† . (4.10)
0 =1
(4.10) can also be expressed as

R m A r = A r Dm (4.11)
where the columns of Ar and diagonal elements of the diagonal matrix Dm (4.2) are,
by definition, respectively the eigenvectors and eigenvalues of Rm . Let u(m, p) and
β (m, p) be respectively the pth eigenvector and eigenvalue of Rm (β (m, p) is one of the
P largest eigenvalues). Further, we can write u(m, p) = α(m, p) e j φ(m, p) ar p and β (m, p) =
ae(m, p) where α(m, p) and φ(m, p) respectively represent the scaling factor and phase shift
introduced during the eigendecomposition. Since the eigendecomposition of the M − 1
matrices Rm (with m 6= mo ) is performed independently, therefore an additional step is
required to pair all the M − 1 sets of the eigenvalues and eigenvectors. Classically, the
pairing can be done by comparing the inner product of the eigenvectors from the fact
that the inner product of two aligned vectors is greater than that of two nonaligned
vectors [1].
ρr p and θr p can be estimated from the angular part of u(m, p) . The argument of a
complex number can only be calculated in its principal form. Therefore, phase un-
wrapping is necessary [115]. Phase unwrapping also introduces phase shift. As we
know that arg{ ar( p, no ) } = 0, therefore the unwrapped phase at n = no is the phase shift
introduced during the eigendecomposition as well as unwrapping. Thus, subtracting
this phase shift from all the remaining components of the unwrapped phase vector
gives an estimation of the true phases. The eigendecompositions of the M − 1 Rm ma-

trices provide M − 1 eigenvectors associated to the signal reflected by the pth target.
Thus, M − 1 estimates of the true phases can be obtained for each target. At high SNR
each estimation will be the same, however, when SNR is low, it is better to combine
them by averaging. Let δ̂r( p, n) be the estimated value of δr( p, n) obtained from the av-
eraged estimate of the true phase by dividing it by −2 π/λ. The estimated values of
the range from the receiving array (ρ̂r p ) and DOA (θ̂r p ) can be retrieved from δ̂r( p, n) by
using the procedure given in section 3.3.2.
− j 2 π δe /λ
The argument of the eigenvalue β (m, p) = e (m, p) can be used to estimate
the transmitter side ranges and DODs of the targets. Being a complex entity, the un-
wrapped phase of β (m, p) should be estimated, which is followed by the phase shifting
with respect to the phase of reference component like before to provide the true un-
wrapped phase vector. Since, (1.4) and (1.5) have the same expression, the similar
steps are conducted as in the case of the estimation of the receiver side location param-
eters. The estimated value of δe(m, p) is obtained from the true unwrapped phase and
then used to build a system of linear equations whose mth equation is given by (3.25)
in section 3.3.2. The estimated values of the range from the transmitting array (ρ̂e p )
and DOD (θ̂e p ) can be retrieved from δ̂e(m, p) by using the method used in section 3.3.2.
Finally, t̂ p can be estimated from ρ̂e p , θ̂e p , ρ̂r p , and θ̂r p with the help of the method given
in section 3.4. All the steps to implement this method is summarize in Algorithm 4.1.
Algorithm 4.1 Submatrix Based Method.

1: Estimate the covariance matrix R from the L samples of the received signal y(t).
2: Find the eigenvectors and eigenvalues of R and compute the mean of MN − P
smallest eigenvalues.
3: Subtract the above mean from the P largest eigenvalues and compute the reduced
noise covariance matrix R̆ as mentioned in (4.5).
4: Break R̆ into M2 submatrices R̆(m, m0 ) of size N × N.
5: Calculate Rm for each m ∈ {1, 2, . . . , M} using the submatrices as given in (4.10).
6: Find the eigenvectors and eigenvalues of Rm for all m and pair them using the
inner product of the eigenvectors for different pairs of m.
7: Compute the principal values of the components of each eigenvector paired above
and unwrap the phases followed by their normalization with respect to the refer-
ence component of each eigenvector.
8: Divide the mean of all the unwrapped and normalized phases calculated above by
−2 π/λ to get δ̂r( p, n) .
9: Obtain ρ̂r p and θ̂r p by applying TLS on (3.28) derived from (3.27) using δ̂r( p, n) .
10: Unwrap, normalize, and divide by −2 π/λ the principal values of the components
of the paired eigenvalues along m for all m to get δ̂e(m, p) in order to obtain ρ̂e p and
θ̂e p by using TLS on (3.26).
11: Finally, use the method given in section 3.4 to obtain t̂ p from ρ̂e p , θ̂e p , ρ̂r p , and θ̂r p .
4.2. PARAFAC Based Method 85
4.2 PARAFAC Based Method

For the L time samples, y(t) can be rewritten as [112]
YM = ( A e Ar ) S T + WM ∈ C MN × L (4.12)
where S ∈ C L× P consists of s(t) at L different time instances. Every element of YM in

(4.12) is associated with three parameters related, respectively, to the receiving antenna,
transmitting antenna, and time sample. Therefore, YM can be rearranged like a three-
way tensor Y ∈ C N × M× L , as shown in Figure 4.1. Creating a tensor out of lower
dimensional data is known as tensorization [120].
N 1
N 2
L
⇒ Y
N
..
.
M
N M
YM
Figure 4.1 – Tensorization.
PARAFAC decomposition of tensor Y is used to get the estimates of Ar , Ae , and

S matrices [96]. Tensor operations are usually performed in its equivalent matrix
form [65,66,96]. The process of creating a matrix out of a tensor is known as matriciza-
tion [120]. Like YM , Y can be matricized into the following two additional matrices
YL = (S Ae ) ArT + WL ∈ LM × N
(4.13)
and
YN = ( Ar S) AeT + WN ∈ NL× M
. (4.14)
According to the least squares principle, the following objective functions can be
written from (4.12), (4.13), and (4.14)

2
T
Ŝ = arg min YM − ( Ae Ar ) S , (4.15)
S F
2

Âr = arg min YL − (S Ae ) ArT , (4.16)
Ar F
and 2

Âe = arg min YN − ( Ar S) AeT (4.17)
Ae F
where Âr , Âe , and Ŝ denote the estimated values of Ar , Ae , and S respectively.
The TALS algorithm is a classical method to minimize the above objective func-
tions [65, 66, 96]. Least squares estimates of (4.15)–(4.17) are given by
h iT
†
Ŝ = ( Ae A r ) YM , (4.18)
h iT
†
Âr = (S Ae ) YL , (4.19)
and h iT
†
Âe = ( Ar S) YN . (4.20)
In the TALS algorithm, (4.18)–(4.20) are alternatively updated with the new val-
2
ues of Âr , Âe , and Ŝ until a stopping criteria is met. kYM − ( Ae Ar ) S T F < etol
is often used as the stopping condition, where etol is the tolerance. In practice, the
algorithm given in [78] is used for PARAFAC decomposition, which uses compres-
sion, line search, normalization, etc. to accelerate the TALS method. The parameter
identifiability of the PARAFAC decomposition is well studied in the existing litera-
ture [69, 70, 81, 96]. It is clearly mentioned in [96] that the PARAFAC decomposition is
unique when max( M, N ) ≥ 3 and P ≤ M N − min( M, N ) for the non-coherent targets
and Swerling model II which are fully satisfied by our assumptions.
According to [69], the matrices obtained by PARAFAC decomposition of a three-

way tensor are scaled and permuted. The permutation has no impact because the
columns of the matrices are paired. However, in the proposed method, the scaling
factor must be removed by dividing all the elements of the directional vectors with
their corresponding reference elements.
Then, a direct approach to estimate t p could be the minimization of the following

cost function
2 2

t̂ p = arg min âe p / âe(mo , p) − ae (t p ) + âr p / âr( p, no ) − ar (t p ) (4.21)
tp F F
4.2. PARAFAC Based Method 87
where âe p and âr p are the estimated directional vectors obtained from the PARAFAC
decomposition, and âe(mo , p) and âr( p, no ) are their respective reference elements used here
to remove the scaling factor in the decomposition of the received signal tensor.
Even though a near-field region occupies a finite space, minimizing (4.21) by using
grid search or Newton’s method is computationally expensive. Therefore, we choose
an indirect method in which we estimate the ranges and directional angles, followed by
the estimation of the coordinates. Let δ̂e(m, p) and δ̂r( p, n) be the estimated path differences
which are obtained from the estimated directional vectors as follows
h n o n oi
δ̂e(m, p) = −λ arg âe(m, p) − arg âe(mo , p) / (2 π ) (4.22)
and h n o n oi
δ̂r( p, n) = −λ arg âr( p, n) − arg âr( p, no ) /(2 π ) (4.23)
where arg {•} represents the unwrapped phase value i.e. the estimated argument [115].
Equations (4.22) and (4.23) can be described as follows:
1. Get the directional vectors âe p and âr p from the PARAFAC decomposition.
2. Extract the arguments of all the components of âe p and âr p .
3. Unwrap the phase vectors obtained from Step 2.
4. Subtract the unwrapped phase corresponding to âe(0, p) and âr( p, 0) from all the com-
ponents of the unwrapped phase vector of âe p and âr p , respectively.
5. Divide each component of the normalized phase vectors obtained from the above
step by −2 π/λ to get δ̂e(m, p) and δ̂r( p, n) .
Once we have δ̂e(m, p) and δ̂r( p, n) , the procedure in section 3.3.2 can be used to obtain
ρ̂e p , θ̂e p , ρ̂r p , and θ̂r p . Finally, t̂ p can be found by using (3.31) which is computation-
ally less expensive and reliable than (4.21). Algorithm 4.2 provides a summary of this
proposed method.
Algorithm 4.2 Tensor Decomposition Based Method.

1: Construct the three-way tensor Y from the received data.
2: Estimate Ae and Ar from Y using PARAFAC decomposition.
3: Use (4.22) and (4.23) to obtain δ̂e(m, p) and δ̂r( p, n) from the estimated Ae and Ar , re-
spectively.
4: Create the system of linear equations by substituting δ̂e(m, p) and δ̂r( p, n) in (3.25) and
(3.27), respectively, for all the values of m and n for each target.
5: Separately solve each system of linear equations created in step 4 using the total
least squares technique to obtain ρ̂e p , θ̂e p , ρ̂r p , and θ̂r p .
6: Substitute the four estimated location parameters in (1.37) and (1.38) to obtain the
parametric equations of the circles, and minimize (3.31) to estimate ϕ̂e and ϕ̂r .
7: Finally, substitute ϕ̂e and ϕ̂r along with ρ̂e p , θ̂e p , ρ̂r p , and θ̂r p in (1.37) and (1.38) to
get the estimated coordinates t̂ p of the pth target.
Consider a bistatic MIMO system as shown in Figure 1.11 with M = 8, N = 9,

mo = 1, no = 1 and de = dr = λ/2. The values of eo, ro, dce and dcr are kept the same as
in section 3.5. Here, we use the accurate signal model in terms of Cartesian coordinates
of the targets, therefore, we can use (3.32) to obtain the RMSE in the estimated position.
Figure 4.2 shows the RMSE in the estimated positions using the proposed methods
elaborated in this chapter. To plot this figure, we have used L = 103 data samples,
K = 103 Monte Carlo iterations. The position vectors of the two chosen near field
targets are [2λ, 7λ, 9λ] T and [3λ, −5λ, −8λ] T .
In Figure 4.2, we can see that the two proposed accurate model based methods have
almost the same performance. Precisely, the PARAFAC decomposition based method
has better performance than the proposed method in section 4.1. One of the reasons is
that PARAFAC exploits all the available information together unlike the other method
which uses sub-covariance matrices and process the information in chunks.
The proposed accurate model based methods is compared in Figure 4.3 with the
proposed approximated model based method by using the same simulation parame-
ters as used in section 3.5. From the figure, we can observe that at low SNR the approx-
imated model based method surpasses the accurate model based methods. However,
because of the Fresnel approximation, at high SNR the accurate model based methods
show better performance.
4.4. Conclusion 89
0
10
ǫ(t1 ); Section 4.1
−1
10 ε(t2 ); Section 1.5
RMSE (λ)
−2
10
−3
10
0 5 10 15 20 25 30
SNR (dB)
Figure 4.2 – RMSE in the position estimated by the proposed methods described in
section 4.1 and section 4.2 versus SNR. de = dr = λ/2, K = 103 , L = 103 , M = 8,
N = 9, mo = 1, no = 1, and P = 2. The CRLBs are calculated using (1.43) and (1.45).
4.4 Conclusion
From the simulation results it is clear that the accurate model based methods can
easily surpass an extremely good approximated model based method because of the
presence of bias due to the Fresnel approximation. Among the proposed accurate
model based methods, the PARAFAC decomposition based method provides better
estimates than the sub-covariance matrices based method at the cost of computation
time because PARAFAC uses iterative TALS based method. The other advantage of an
accurate model methods is that it can be used for a system which uses half wavelength
inter element spacing to minimize mutual coupling between the array elements.
0
10
−1
10
RMSE (λ)

−2
10 ǫ(t1 ); Section 4.1
−3
10
0 5 10 15 20 25 30
SNR (dB)
Figure 4.3 – RMSE in the position estimated by the proposed methods described in
section 2.2, section 4.1, and section 4.2 versus SNR. de = dr = λ/4, K = 103 , L = 103 ,
M = 7, N = 9, mo = 4, no = 5, and P = 2. The CRLBs are calculated using (1.43) and
(1.45).
Conclusion
1 Conclusions and Perspectives

The primary objective of the work in this thesis is to improve the localization per-
formance of the targets in the near field regions of the transmitting and receiving arrays
of a bistatic MIMO system. This thesis contains several proposed methods in this con-
text for two different signal models with and without Fresnel approximation of the
spherical wavefront.
The proposed methods are dedicated to the near field region of a bistatic MIMO
radar with ULAs. Fresnel approximation is usually opted to locate the sources in the
near field region of an array. Thus, we have proposed a localization method and com-
pared it with an existing method and CRLB. The simulation results show that the per-
formance of the proposed method is better in terms of RMSE and maximum number
of localizable targets than the existing method. A drawback of the proposed method
is its computational complexity. A flaw in the analysis of this proposed method is that
the signal model used in its analysis is based on Fresnel approximation. However, for
a point like target the wavefront is spherical. Hence, we can say that the analysis of an
approximated model based method with an approximated model based signal data is
biased.
When an approximated model based method is analyzed with the spherical wave-
front based signal model, the location parameters show some biases. This bias can be
considered as a function of the true location parameters and array parameters. If the
array parameters, like number of antenna and inter element distance, are constant, the
bias will be constant throughout the near field region. This property of the bias allows
us to put it in the category of systematic error. This bais stands out at high SNR because
at low SNR the additive noise dominates. At different locations, the bias has different
values, therefore, the SNR at which the bias dominates the noise may vary with the
location. We have proposed two methods to reduce the effect of the biases from lo-
cation parameters. The first method is based on LUT. This proposed method can be
subdivided into two steps. In the first step, we create the LUTs for the transmitting
and receiving arrays. And in the second step, we reduce the biases in the location pa-
rameters by interpolation. This approach is effective, however, it is slow because find-
91
92 Conclusion
ing four nearest neighbors from two LUTs is time consuming. To overcome the space
and time requirement of LUT based correction method, we have proposed the second
method which estimates the accurate relative propagation path length vector and cor-
rects the location parameters using TLS. The proposed methods add extra complexity
to the localization process however greatly improves the estimation performance.
Beside the methods to reduce the effects of Fresnel approximation, we have pro-
posed two methods which can directly process the spherical wavefront based signal
model. The first method is based on subspaces whereas the second method is based
on tensor decomposition. Principally in both the proposed methods, we estimate the
directional vectors first and then the location parameters from them. Tensor decom-
position usually uses iterative methods to find its components which makes it com-
putationally costly. On the other hand, subspaces based proposed method processes
submatrices which is faster than the tensor decomposition based method however the
maximum number of localizable targets reduces to the number of receiving antennas.
Because of the complete exploitation of the signal data all together by the tensor de-
composition based proposed method, its performance surpasses the proposed sub-
space based method. Using the accurate model doesn’t imply that the performance in
terms of RMSE will always be better. This can be observed from the simulation results
that if the proposed correction methods are applied on the proposed approximated
model based method, it can surpass both the accurate model based methods. How-
ever, it is clear from this work that approximated model alone is not a great choice
when we are looking for precision. It is better to correct the bias or to use the accurate
model based methods in the near field region of a bistatic MIMO system. Additionally,
the proposed accurate model based methods support interelement distance between
λ/4 and λ/2 which can not be handled by most of the approximated model based
methods.
2 Future Work
The performance of all the proposed methods in this thesis is analyzed only by
computer simulations. However, the methods are proposed for real life applications.
Therefore, the analysis is incomplete without testing the proposed methods with some
real bistatic MIMO radar data.
In the near field region of an array, the fast moving target may not induce coherent
Doppler shift. Thus, the model used in this study may not be suitable for such targets
because in this study we have ignored the effects of Doppler shifts in the signal model.
In the literature, we can find countless work on the point-like sources in the near field
region of an array. However, we may come across large targets which can not be local-
ized properly with the proposed methods. Thus, this work should be extended to the
2. Future Work 93
fast moving distributed targets.

To handle the spherical wavefront based signal model, we have proposed to esti-
mate directional vectors first and then extract the location parameters from them. The
directional vectors are always in the complex form, therefore, the phases of these com-
plex components can only be retrieved in their principal form. To extract the location
parameter, we need unwrapped phases. There exists a working algorithm to unwrap
theses principal valued phases. However, due to the estimation errors in the estimated
directional vectors, the existing unwrapping algorithm induces a large error which can
not be ignored. Thus, further research can be done to find an algorithm which can un-
wrap the phases without error.
The method in section 4.1 uses denoised covariance matrix. The used denoising
method is not a best approach. Also, the advanced covariance estimators are not used.
Therefore, the estimation accuracy of the proposed method in section 4.1 can further
be improved by using an advanced covariance estimators and a better denoising tech-
nique.
The proposed approximated model based method in section 2.2 has high estima-
tion accuracy, however it is slow because of the spectral search. If we try to convert
the problem into a polynomial one, then we will get a sparse high order polynomial
whose roots finding time will be very large because of the eigen decomposition in
its background. Therefore, instead of decreasing, the computation time will increase.
However, if we use a fast root finding algorithm which only looks for the P closest
roots to the unity circle, then the polynomial rooting can be a useful approach.
A
Biased Location Parameters
As mentioned before, the propagation of the systematic error introduced by the

Fresnel approximation may lead to different bias in the same location parameters. In
the following, we show different possible approaches to obtain biased location param-
eters and the distribution of the bias attached to them in the Fresnel region of ULA.
Like in section 1.3.2, we use the receiving array parameters here to elaborate the error
distribution. In order to calculate the biased location parameters corresponding to the
receiving array, we choose a location (ρr p , θr p ) in its near field region and construct
δr (ρr p , θr p ) vector and extract the range and DOA from it using the methods given be-
low. The same steps can be used for the transmitting ULA. To calculate the biased
location parameters, there is not need to construct the directional vectors.
A.1 Biased Location Parameters Extraction Methods
A.1.1 Vector Flip

This biased location parameters extraction method is inspired by the method in [1],
therefore it requires symmetric array described in section 1.3.3.
From (1.32), we can write
T
JN δer p − δer p = 2 ωr p − N̆, − N̆ + 1, . . . , 0, . . . , N̆ − 1, N̆ (A.1)
and h iT
JN δer p + δer p = 2 φr p N̆ 2 , ( N̆ − 1)2 , . . . , 0, . . . , ( N̆ − 1)2 , N̆ 2 . (A.2)
95
96 Appendix A. Biased Location Parameters
Using the similar operations on δr p , we can write
δ̀r p = JN δr p − δr p (A.3)
and
δ́r p = JN δr p + δr p . (A.4)
Further, δ̀r p can be expanded as

     
− N̆ γr( p, N ) − γr( p, 1) − N̆
     
− N̆ + 1  γr( p, N −1) − γr( p, 2)  − N̆ + 1
     
 .
.
  ..   .
.

 .   .   . 
     
     
 −1  γr( p, no +1) − γr( p, no −1)   −1 
     
δ̀r p = 2 ωr p 
 0  + ρr 
 p  0  ≈ 2ω
 e r p
 0 
  (A.5)
     
 1  γr( p, no −1) − γr( p, no +1)   1 
     
 ..   ..   .. 

 . 


 . 


 . 

 N̆ − 1     N̆ − 1 
   γr( p, 2) − γr( p, N −1)   
N̆ γr( p, 1) − γr( p, N ) N̆
From (A.5), ω
e r p can be estimated as
h i h i
1 2 N̆ ρr p
erp
ω =
4 N̆
∑ δ̀r p
i +1
− δ̀r p
i
= ωr p +
2 N̆
γr( p, 1) − γr( p, N ) (A.6)
i =1
which can be used to obtain the biased DOA using the following expression
θer p = arccos(ω
e r p /dr ). (A.7)
As above, expanding δ́r p gives

     
− N̆ 2 γr( p, N ) + γr( p, 1) − N̆ 2
     
( N̆ − 1)2   γr( p, N −1) + γr( p, 2)  ( N̆ − 1)2 
     
 ..   ..   .. 
 .   .   . 
     
     
 1  γr( p, no +1) + γr( p, no −1)   1 
     

δ́r p = 2 φr p  0  + ρ r
 0  ≈ 2 e
φr
 0 , (A.8)
 p   p  
     
 1  γr( p, no −1) + γr( p, no +1)   1 
     
 ..   ..   .. 

 . 


 . 


 . 

( N̆ − 1)2    ( N̆ − 1)2 
  γ
 r( p, 2) + γr( p, N −1)   
N̆ 2 γr( p, 1) + γr( p, N ) N̆ 2
A.1. Biased Location Parameters Extraction Methods 97
er p can be calculated as follows

from which φ
N̆ h i h i
er p = 1
φ ∑ i2
1
δ́r p + δ́r p (A.9)
4 N̆ i =1
N̆ −i +1 N̆ +i +1
ρr p N̆ γr( p, no +i) + γr( p, no −i)

= φr p +
2 N̆
∑ i2
(A.10)
i =1
where no = N̆ + 1. Finally, the biased range can be calculated as
ρer p = d2r sin2 (θer p )/(2 φ

er p ). (A.11)
A.1.2 Double Difference

It can be observed in (1.25) that the coefficients of ωr p is linear. Therefore, taking
difference between the consecutive elements of δer p make the coefficients of ωr p constant
which can be eliminated by the second difference. This procedure also puts a constraint
of N ≥ 3 on this extraction method.
Difference between the consecutive elements of a vector of length N can be obtained
by multiplying that vector with the following matrix
 
−1 1 0 ··· 0 0
 
 0 −1 1 ··· 0 0
 
 . .. .. .. .. .. 
PN =  .. . . . . . ∈ R( N −1)× N . (A.12)
 
 
 0 0 0 ··· 1 0
0 0 0 ··· −1 1
First difference of δr p can be written as

   
−ωr p + [2(1 − no ) + 1] φr p γr( p, 2) − γr( p, 1)
   
 −ωr p + [2(2 − no ) + 1] φr p   γr − γr( p, 2) 
   ( p, 3) 
PN δr p =  ..  + ρr p  .  (A.13)
 .   .. 
   
−ωr p + [2( N − 1 − no ) + 1] φr p γr( p, N ) − γr( p, N −1)
 
−ω e r p + [2(1 − n o ) + 1] φ er p
 
 −ω e r + [ 2 ( 2 − n o ) + 1 ] e
φr 
 p p 
≈ ..  ∈ R N −1 (A.14)
 . 
 
e r p + [2( N − 1 − n o ) + 1] φ
−ω er p
because (n + 1 − no )2 − (n − no )2 = 2 (n − no ) + 1.
Similarly, on multiplying (A.14) by PN −1 ∈ R( N −2)×( N −1) , we can obtain the second
difference as

  
2 φr p γr( p, 3) − 2 γr( p, 2) + γr( p, 1)
   
2 φr p   γr − 2 γr + γr 
   ( p, 4) ( p, 3) ( p, 2) 
PN −1 PN δr p =  .  + ρr p  ..  (A.15)
 ..   . 
   
2 φr p γr( p, N ) − 2 γr( p, N −1) + γr( p, N −2)
 
2φer p
 
2 φer p 
 
≈  .  ∈ R N −2 . (A.16)
 .. 
 
e
2 φr p
er p can be calculated by the half of the mean of the elements of PN −1 PN δr p as

The φ
N −2 h i
1
er p =
φ
2 ( N − 2) ∑ PN −1 PN δr p
i
(A.17)
i =1
ρr p N −2
= φr p +
2 ( N − 2) ∑ γr( p, i+2) − 2 γr( p, i+1) + γr( p, i) . (A.18)
i =1
er p can be used with (A.14) to get ω

Furthermore, φ e r p as
N −1 h i
1
erp =
ω
N−1 ∑ [2 (i − no ) + 1] φer p − PN δr p
i
(A.19)
i =1
" #
γr( p, 1) − γr( p, N ) N −2
( N − 2 n o + 1)
= ωr p + ρ r p
N−1
+
2 ( N − 2) ∑ γr( p, i+2) − 2 γr( p, i+1) + γr( p, i) .
i =1
(A.20)
Finally, θer p and ρer p can be calculated from ω er p by using (A.7) and (A.11)
e r p and φ
respectively.
A.1.3 Overdetermined System of Linear Equations
For N ≥ 3, (1.34) can be considered an overdetermined system of linear equations

with ωr p and φr p as its two unknowns. This system of linear equations can be solved by
using conventional LS estimator. Applying this concept on δr p can provide the biased
location parameters as
" # " #
erp
ω † ωr p
= N δr p = + ρr p N † γr p (A.21)
e
φr p φr p
A.1. Biased Location Parameters Extraction Methods 99
where γr p = [γr( p, 1) , γr( p, 2) , · · · , γr( p, N ) ] T . As before, θer p and ρer p can be calculated from
er p by using (A.7) and (A.11) respectively.
e r p and φ
ω
A.1.4 Least Squares Error
Like the above method in section A.1.3, this approach is also based on LS estima-
tion. However, it directly estimates biased range and DOA by minimizing the follow-
ing cost function
2
e e
(ρer , θr ) = arg min δr − δr (ρ, θ )
p p p (A.22)
(ρ, θ ) F
with the help of a numerical method like Newton’s minimization technique because
closed form expression may not be easily found. The use of numerical methods to
minimize the cost function makes this method slow because of the iterations; and un-
reliable due to the local minimums.
A.1.5 Vector Component Pairs
A pair of rows of δer p makes a perfect system of linear equations in ωr p and φr p .

Using this fact, ω er p can be calculated from δr p .
e r p and φ
The ith and i0 th components of δr p can be used as follows
" # " #" #
erp
ω 1 (i0 − no )2 −(i − no )2 δr( p, i)
= . (A.23)
er p
φ (i − i0 )(i − no )(i0 − no ) (i0 − no ) −(i − no ) δr( p, i0 )
Estimations from all the possible combinations of i and i0 can be combined using mean
which gives
2 ρr p N −1 N (i0 − no )2 γr( p, i) − (i − no )2 γr( p, i0 )

erp
ω = ωr p +
( N − 2)( N − 1) ∑ ∑ (i − i0 )(i − no )(i0 − no )
(A.24)
i =1 i 0 = i +1
i 6=no i 0 6=no
and
2 ρr p N −1 N (i0 − no ) γr( p, i) − (i − no ) γr( p, i0 )

er p = φr p +
φ
( N − 2)( N − 1) ∑ 0∑ (i − i0 )(i − no )(i0 − no )
. (A.25)
i =1 i = i +1
i 6=no i 0 6=no
As above, θer p and ρer p can be calculated by using (A.7) and (A.11) respectively.
A.1.6 Residue Vector Approximation

The residue can be approximated by the third order term of (1.12) as
(n − no )3 d3r
2
γr( p, n) ≈ cos θ rp sin θr p (A.26)
2 ρ3r p
which can be used as

" # " #
erp
ω ωr p
= + ρr p N † γr p (A.27)
e
φr p φr p
 
(1 − n o )3
 
 (2 − n o )3 
" #  
 ..  3
ωr p †
 .  dr
 2
≈
φr p
+N 
0  2 ρ2 cos θr p sin θr p (A.28)
  rp
 . 
 . 
 . 
( N − n o )3
to obtain the biased range and DOA by using (A.11) and (A.7) respectively.
A.2 Numerical Example

In section A.1, we can see that different approaches lead to different biases in the
location parameter. For a linear array, the range and DOA cover a 2D space. Thus,
for easy visualization of the error distribution in the near field region, the Fresnel ap-
proximation error between the true and biased range and DOA is calculated in terms
of Euclidean distance as
r
2 2
e
ξrp = e
ρer p cos θr p − ρr p cos θr p e
+ ρer p sin θr p − ρr p sin θr p
r
2 2 e
= ρer p + ρr p − 2 ρer p ρr p cos θr p − θr p . (A.29)
Figure A.1, Figure A.2, Figure A.3, Figure A.4, Figure A.5, and Figure A.6 show the pro-
file of this error (ξer p ) for the biased location parameters extraction approaches briefed
in section A.1.1, section A.1.2, section A.1.3, section A.1.4, section A.1.5, section A.1.6
respectively. The symmetric ULA used in this example is assumed to have N = 9 an-
tennas and dr = λ/4 inter element spacing. In the figures, the lower and upper limits
of the Fresnel region are also drawn to show the near field region of the considered
array.
From the error profiles of these methods, we can notice that the biases in the lo-
A.2. Numerical Example 101
cation parameters due to the Fresnel approximation have significant variation with
respect to DOA except for the method in section A.1.3. Additionally, as shown in Fig-
ure A.3, the error magnitude is also small and gradually decreases with the increase
in range which is close to ideal behavior of a function inversely proportional to the
range. Because of its closeness to the ideal error profile, we have used this method to
construct LUT in section 3.2. And it has shown acceptable performance in the reduc-
tion of the biases in the location parameters due to the Fresnel approximation which
can be backed up by the numerical examples in section 3.5. On analyzing Figure A.2
visually, we can say that the method in section A.1.2 can be accepted as the second
choice to construct LUTs.
8
ξerp (λ)
0 .0
Perpendicular to the array (λ)
7
25
Fr
6 0.05 0.05 F̆r
5
4 0.0
75
3 0 75
0.
0.0 0.07
5 5
2
0. 0 .1
0.1
1 1
0.
0 .3 0.3
2
2
0.
0
−8 −6 −4 −2 0 2 4 6 8
Along the array (λ)
Figure A.1 – Approximation error (ξer p ) profile of the extraction method in section A.1.1;
dr = λ/4, N = 9, and no = 5.
8
0.025 0.05 0.025
7
ξerp (λ)
6
Fr
5 7 5
0.0 F̆r
4 0.05 0.05
0.1
3
5 0.075
2 0.07
0.
1
1
1
3
0.
2 0.
0.
0
−8 −6 −4 −2 0 2 4 6 8
dr = λ/4, N = 9, and no = 5.
8
0.0
2 5 25
0.0
7 ξerp (λ)
6 Fr
5 F̆r
4 0 .0
5
3
5
0 .0
5
0.02
2 0.1
0.02
0.07
0 .0 7 5
5
1 0.2
5
0.3
0
−8 −6 −4 −2 0 2 4 6 8
dr = λ/4, N = 9, and no = 5.
A.2. Numerical Example 103
0.02
7
5
ξerp (λ)
0.02
5
6 Fr
1
0 .0
0.0
1
5 F̆r
3
0.01
0.01
0.05
2
0.1
0.0
0. 0
0.025
25
1 25 0.07
5
5.
0.0 0.07 00.2 50
0.05
3
0 .0
.0
0.01 25 0.01
0
5
−8 −6 −4 −2 0 2 4 6 8
dr = λ/4, N = 9, and no = 5.
8
0. ξerp (λ)
02
7 5
5 02F
0.025
0. r
0.025
6
0.0
F̆r
0.01
1
5 0.0
1
5
0.0
5
0.01
4 0.0
0.075 0.075
3
0.0
0 .1
1 0.0.05
25
25
2 1
0. 0.
0
1
0.05
0.2 .2
5 30
0.0
1 0.3
0.
0.20.07 0.
75
0
−8 −6 −4 −2 0 2 4 6 8
dr = λ/4, N = 9, and no = 5.
8
ξerp (λ)
7
0.
02
5
Fr
02
5
6
0.
0.05 F̆r
0.01
5 5
0 .0
0 .0 1
4 0.075 0.075
3 0.1
0.1
2
00.0.000.05.1
0.2
0.2
2575
1 0.
3
0
−8 −6 −4 −2 0 2 4 6 8
dr = λ/4, N = 9, and no = 5.
B
Correction Methods for Single Array
The two proposed correction methods in chapter 3 can be used for a single receiv-
ing array [105, 106]. In the following, we apply them to correct the biased location pa-
rameters estimated by the two existing approximated model based near field sources
location methods in [2] and [3]. Consider a ULA with N = 9 receiving antennas and
d = λ/4 inter element spacing. Since the methods in [2] and [3] require symmetric
array, we take no = 5 which makes N̆ = 4. The received signals for the simulations are
generated using the model based on spherical wavefront. The actual location param-
eters of the pth location is given by (ρr p , θr p ). As in Appendix A, (ρer p , θer p ) denotes the
pair of biased location parameters estimated here by using [2] and [3]. And (ρ̂r p , θ̂r p )
denotes the corresponding pair of corrected location parameters. The estimation er-
ror (ξer p ) between the actual and biased location parameters is calculated in terms of
Euclidean distance using the expression in (A.29). By using the similar expression,
we can calculate the estimation error (ξ̂ r p ) between the actual and corrected location
parameters as r
ξ̂ r p = ρ̂2r p + ρ2r p − 2 ρ̂r p ρr p cos θ̂r p − θr p . (B.1)
Figure B.1 and Figure B.2 show the contour plots of the Euclidean distances be-
tween the actual and biased locations estimated in noiseless situation by the methods
in [2] and [3] respectively. The contours are generated by discretizing the near field re-
gion into a regular rectangular grid and then calculating ξer p at each of these locations.
Each location is handled separately, therefore the methods in [2] and [3] experience a
single source situation. The upper and lower limits of the Fresnel region are also shown
along with the contours to easily identify the near field region. L = 103 samples are
105
106 Appendix B. Correction Methods for Single Array
used to estimate the biased location parameters from both the methods to avoid the
degradation in their performance due to the lack of samples, because both are based
on subspace principle. The unit of the approximation error values shown on the con-
tour lines is λ. These figures have similar meaning as the figures in section A.2. They
also conclude that two different methods have different approximation error profiles.
8
ξerp (λ)
7
Fr
6 0.04
F̆r
0.03
5
0.03 6
4 0.0
0.04 0.04
3
0.1 0.06
2 6
0.0 0.08 0.08
0.1
1 0.1
0
−8 −6 −4 −2 0 2 4 6 8
Figure B.1 – ξer p of [2] in noiseless condition; d = λ/4, L = 103 , and N̆ = 4.
8
ξerp (λ)
7
Fr
6
F̆r
0.0
5
3
4
0 .0
3
0.0
0.0
4
3
0. 0
0.1
6
2 0.06
8 0.08
1 0 . 0 .1 0.1
0.04
0
0
−8 −6 −4 −2 0 2 4 6 8
Figure B.2 – ξer p of [3] in noiseless condition; d = λ/4, L = 103 , and N̆ = 4.
Let ξ r p = ξer p /ξ̂ r p be the ratio of the error in the location parameters at the pth lo-
cation before correction to the error after correction. Figure B.3 and Figure B.4 respec-
tively show this ratio for the near field region when the biased location parameters of
107
each location are estimated in noiseless condition by using the methods in [2] and [3]
and corrected by our proposed method in section 3.2. The LUT is constructed by us-
ing 100 uniformly spaced discrete values of the ranges between F̆r and Fr ; and 720
uniformly spaced discrete values of the DOA from 0.1◦ to 179.9◦ .
8
8
5
8
6
43
2
8
4
1.
68
34
5
2
3
5
ξrp
1.
66
2 Fr 6
F̆r
1
1.5 1.5
0
−8 −6 −4 −2 0 2 4 6 8
Figure B.3 – ξ r p of [2], corrected using section 3.2; d = λ/4, L = 103 , and N̆ = 4.
8
6
8
5
86
4
4
2.
3
5
8
46
3 2 5
2.
3
ξrp 6 2
2 6
Fr
1
F̆r
0
−8 −6 −4 −2 0 2 4 6 8
Same as above, Figure B.5 and Figure B.6 respectively show the profile of ξ r p in the
near field region when the biased location parameters of each location are estimated in
noiseless condition by using the methods in [2] and [3] and corrected by our proposed
method in section 3.3.
108 Appendix B. Correction Methods for Single Array
8
12 1 2
12 8
8
12
5
43
2
8
4
8
1.
34
5
2
3 5
1.
ξrp
2
Fr
1 .4
1.
1
4
F̆r
0
−8 −6 −4 −2 0 2 4 6 8
8
6
12
5
12
12
4
4
2.
3
12
5
34
3 2 5
2.
ξrp 2
2
Fr
1
F̆r
0
−8 −6 −4 −2 0 2 4 6 8
109
In Figure B.3, Figure B.4 Figure B.5, and Figure B.6, the Fresnel region with ξ r p > 1
signifies that the proposed correction methods reduce the error due the approximated
model. We can observe from these figures that the improvement of the proposed
correction method varies inside the Fresnel region. The principal cause of this phe-
nomenon is that the approximation error due to the higher order terms in (1.14) is
a nonlinear function of range and DOA. In addition, the degree of the nonlinearity
changes inside the Fresnel region. Large value of ξ r p implies better correction. From
the patterns of contours in Figure B.3, Figure B.4 Figure B.5, and Figure B.6, we can
observe that ξ r p does not have the same values in the same part of the near field re-
gion for the two methods. This illustrates that the correction methods have different
performance for different methods.
Bibliography
[1] Erning Zhou, Hong Jiang, and Han Qi. 4-D parameter estimation in bistatic
MIMO radar for near-field target localization. In International Wireless Symposium
(IWS), pages 1–4, Shenzhen, China, March 2015. IEEE. (Cited on pages 12, 15, 16,
31, 47, 51, 52, 53, 55, 57, 61, 62, 63, 64, 81, 83, and 95.)
[2] Wanjun Zhi and Michael Yan-Wah Chia. Near-field source localization via sym-
metric subarrays. IEEE Signal Processing Letters, 14(6):409–412, June 2007. (Cited
on pages 12, 16, 26, 38, 39, 44, 45, 57, 58, 59, 60, 61, 62, 76, 105, 106, 107, and 108.)
[3] Jin He, MNS Swamy, and M Omair Ahmad. Efficient application of MUSIC algo-
rithm under the coexistence of far-field and near-field sources. IEEE Transactions
on Signal Processing, 60(4):2066–2070, April 2012. (Cited on pages 16, 26, 44, 45,
61, 76, 105, 106, 107, and 108.)
[4] Merrill Ivan Skolnik. Introduction to Radar Systems. Electrical engineering series.
McGraw-Hill, 2001. (Cited on pages 24 and 36.)
[5] Merrill Ivan Skolnik. Radar Handbook, Third Edition. Electronics electrical engi-
neering. McGraw-Hill Education, 2008. (Cited on pages 24 and 35.)
[6] David J. Daniels. Ground Penetration Radar. Radar. The institution of Electrical
Engineers, 2004. (Cited on page 24.)
[7] Kai Luo. MIMO RADAR: Target Localization. PhD thesis, Imperial College, Lon-
don, UK, March 2011. (Cited on page 24.)
[8] Jack Capon. High-resolution frequency-wavenumber spectrum analysis. Pro-
ceedings of the IEEE, 57(8):1408–1418, August 1969. (Cited on pages 25 and 28.)
[9] Harry L. Van Trees. Optimum Array Processing: Part IV of Detection, Estimation,
and Modulation theory. John Wiley & Sons, 2004. (Cited on pages 25, 28, and 29.)
[10] V. F. Pisarenko. The retrieval of harmonics from a covariance function. Geo-
physical Journal of the Royal Astronomical Society, 33(3):347–366, 1973. (Cited on
page 25.)
[11] Georges Bienvenu and Laurent Kopp. Optimality of high resolution array pro-
cessing using the eigensystem approach. IEEE Transactions on Acoustics, Speech
and Signal Processing, 31(5):1235–1248, 1983. (Cited on pages 25 and 28.)
111
112 Bibliography
[12] Ralph O Schmidt. Multiple emitter location and signal parameter estimation.
IEEE Transactions on Antennas and Propagation, 34(3):276–280, March 1986. (Cited
on pages 25 and 28.)
[13] A. Barabell. Improving the resolution performance of eigenstructure-based
direction-finding algorithms. In IEEE International Conference on Acoustics, Speech,
and Signal Processing (ICASSP), volume 8, pages 336–339, April 1983. (Cited on
pages 25 and 28.)
[14] Robert Roy, Arogyaswami Paulraj, and Thomas Kailath. ESPRIT–A subspace
rotation approach to estimation of parameters of cisoids in noise. IEEE Transac-
tions on Acoustics, Speech and Signal Processing, 34(5):1340–1342, 1986. (Cited on
pages 25 and 28.)
[15] R Roy, A Paulraj, and T Kailath. Estimation of signal parameters via rotational
invariance techniques-ESPRIT. In 30th Annual Technical Symposium, pages 94–
101. International Society for Optics and Photonics, 1986. (Cited on pages 25
and 28.)
[16] Richard Roy and Thomas Kailath. ESPRIT–Estimation of signal parameters via
rotational invariance techniques. IEEE Transactions on Acoustics, Speech and Signal
Processing, 37(7):984–995, 1989. (Cited on pages 25, 26, and 28.)
[17] Bhaskar D Rao and KVS Hari. Performance analysis of root-MUSIC. IEEE Trans-
actions on Acoustics, Speech and Signal Processing, 37(12):1939–1949, 1989. (Cited
[18] Sylvie Marcos, Alain Marsal, and Messaoud Benidir. The propagator method
for source bearing estimation. Signal processing, 42(2):121–138, 1995. (Cited on
pages 26 and 28.)
[19] Marius Pesavento, Alex B Gershman, and Kon Max Wong. Direction finding in
partly calibrated sensor arrays composed of multiple subarrays. IEEE Transac-
tions on Signal Processing, 50(9):2103–2115, September 2002. (Cited on pages 26
and 28.)
[20] S.S. Reddi. Multiple source location–A digital approach. IEEE Transactions
on Aerospace and Electronic Systems, AES-15(1):95–105, January 1979. (Cited on
pages 26 and 28.)
[21] Donald W Tufts and Ramdas Kumaresan. Estimation of frequencies of multiple
sinusoids: Making linear prediction perform like maximum likelihood. Proceed-
ings of the IEEE, 70(9):975–989, September 1982. (Cited on pages 26 and 28.)
[22] Ramdas Kumaresan and Donald W Tufts. Estimating the angles of arrival of
multiple plane waves. IEEE Transactions on Aerospace and Electronic Systems, AES-
19(1):134–139, January 1983. (Cited on pages 26 and 28.)
Bibliography 113
[23] A. L. Swindlehurst and T. Kailath. Passive direction-of-arrival and range estima-

tion for near-field sources. In Fourth Annual ASSP Workshop on Spectrum Estima-
tion and Modeling, pages 123–128, Minneapolis, Minnesota, August 1988. (Cited
on pages 26, 31, 37, 39, 44, 45, and 61.)
[24] Emmanuèle Grosicki, Karim Abed-Meraim, and Yingbo Hua. A weighted linear
prediction method for near-field source localization. IEEE Transactions on Signal
Processing, 53(10):3651–3660, 2005. (Cited on pages 26, 39, 44, 45, 49, and 61.)
[25] Hongyang He, Yide Wang, and Joseph Saillard. A high resolution method of
source localization in near-field by using focusing technique. In European Sig-
nal Processing Conference, pages 1–5, Lausanne, Switzerland, August 2008. IEEE.
(Cited on pages 26 and 31.)
[26] Yi Duo Guo, Hui Xie, Yong Shun Zhang, Jian Gong, and Di Shen. Localization
for near-field targets based on virtual array of MIMO radar. Radar Science and
Technology, 10(1):82–87, February 2012. (Cited on pages 26 and 31.)
[27] Bo Wang, Yanping Zhao, and Juanjuan Liu. Mixed-order MUSIC algorithm for
localization of far-field and near-field sources. IEEE Signal Processing Letters,
20(4):311–314, April 2013. (Cited on pages 26 and 31.)
[28] Jianzhong Li, Yide Wang, and Wei Gang. Signal reconstruction for near-field
source localisation. IET Signal Processing, 9(3):201–205, 2015. (Cited on pages 26
and 31.)
[29] Gerard J Foschini. Layered space-time architecture for wireless communication

in a fading environment when using multi-element antennas. Bell labs technical
journal, 1(2):41–59, 1996. (Cited on page 27.)
[30] Daniel W Bliss, Keith W Forsythe, Alfred O Hero, and Ali F Yegulalp. En-
vironmental issues for MIMO capacity. IEEE Transactions on Signal Processing,
50(9):2128–2142, September 2002. (Cited on page 27.)
[31] David Gesbert, Helmut Bolcskei, Dhananjay A Gore, and Arogyaswami J

Paulraj. Outdoor MIMO wireless channels: Models and performance prediction.
IEEE Transactions on Communications, 50(12):1926–1934, 2002. (Cited on page 27.)
[32] D.W. Bliss and K.W. Forsythe. Multiple-input multiple-output (MIMO) radar
and imaging: Degrees of freedom and resolution. In Conference Record of the
Thirty-Seventh Asilomar Conference on Signals, Systems and Computers, volume 1,
pages 54–59. IEEE, November 2003. (Cited on pages 27 and 28.)
[33] Eran Fishler, Alex Haimovich, Rick Blum, Dmitry Chizhik, Len Cimini, and
Reinaldo Valenzuela. MIMO radar: An idea whose time has come. In Proceedings
of the IEEE Radar Conference, pages 71–78, April 2004. (Cited on page 27.)
114 Bibliography
[34] Alexander M. Haimovich, Rick S. Blum, and Leonard J. Cimini. MIMO radar
with widely separated antennas. Signal Processing Magazine, IEEE, 25(1):116–129,
2008. (Cited on pages 27 and 36.)
[35] George Turin. An introduction to matched filters. IRE transactions on Information
theory, 6(3):311–329, June 1960. (Cited on page 27.)
[36] Ilya Bekkerman and Joseph Tabrikian. Spatially coded signal model for ac-
tive arrays. In International Conference on Acoustics, Speech, and Signal Processing
(ICASSP), volume 2, pages ii–209–12, Montreal, Que., Canada, May 2004. IEEE.
(Cited on page 27.)
[37] Hugh Griffiths. Multistatic, MIMO and networked radar: the future of radar sen-
sors? In European Radar Conference (EuRAD), pages 81–84, Paris, France, Septem-
ber 2010. IEEE. (Cited on pages 27 and 28.)
[38] Jian Li and Petre Stoica. MIMO radar signal processing. John Wiley & Sons, Inc.,
Hoboken, New Jersey, 2009. (Cited on page 27.)
[39] MC Jackson. The geometry of bistatic radar systems. In IEE Proceedings F (Com-
munications, Radar and Signal Processing), volume 133, pages 604–612. IET, De-
cember 1986. (Cited on page 27.)
[40] Jian Li, Petre Stoica, Luzhou Xu, and William Roberts. On parameter identifia-
bility of MIMO radar. IEEE Signal Processing Letters, 14(12):968–971, December
2007. (Cited on page 27.)
[41] Jian Li and Petre Stoica. MIMO radar with colocated antennas. Signal Processing
Magazine, IEEE, 24(5):106–114, September 2007. (Cited on pages 27, 36, and 44.)
[42] Pier Francesco Sammartino. A comparison of processing approaches for distributed
radar sensing. PhD thesis, University College London, 2009. (Cited on page 27.)
[43] Chen Chen and Xiaofei Zhang. A RD-ESPRIT algorithm for coherent DOA es-
timation in monostatic MIMO radar using a single pulse. International Journal of
Electronics, 101(8):1074–1085, 2014. (Cited on page 28.)
[44] Ping Huanga, Hui Wang, and Cheng Yang. Non circular rootMUSIC algo-
rithm for monostatic MIMO radar. IEICE ELECTRONICS EXPRESS, 11(13), 2014.
(Cited on page 28.)
[45] Wei Wang, Xianpeng Wang, Hongru Song, and Yuehua Ma. Conjugate ESPRIT
for DOA estimation in monostatic MIMO radar. Signal Processing, 93(7):2070–
2075, 2013. (Cited on page 28.)
[46] X Zhang, Y Huang, C Chen, J Li, and D Xu. Reduced-complexity Capon for
direction of arrival estimation in a monostatic multiple-input multiple-output
radar. IET Radar, Sonar & Navigation, 6(8):796–801, 2012. (Cited on page 28.)
Bibliography 115
[47] Xiaofei Zhang and D Xu. Low-complexity ESPRIT-based DOA estimation for
colocated MIMO radar using reduced-dimension transformation. Electronics Let-
ters, 47(4):283–284, 2011. (Cited on page 28.)
[48] Rong Xie, Zheng Liu, and Zi-jing Zhang. DOA estimation for monostatic MIMO
radar using polynomial rooting. Signal processing, 90(12):3284–3288, 2010. (Cited
on page 28.)
[49] Rémy Boyer. Performance bounds and angular resolution limit for the moving
colocated MIMO radar. IEEE Transactions on Signal Processing, 59(4):1539–1552,
April 2011. (Cited on page 28.)
[50] Honglei Miao, Markku Juntti, and Kegen Yu. 2-D unitary ESPRIT based joint
AOA and AOD estimation for MIMO system. In IEEE 17th International Sympo-
sium on Personal, Indoor and Mobile Radio Communications, pages 1–5. IEEE, 2006.
(Cited on page 29.)
[51] Chen Duofang, Chen Baixiao, and Qin Guodong. Angle estimation using ESPRIT
in MIMO radar. Electronics Letters, 44(12):770–771, 2008. (Cited on page 29.)
[52] Chen Jinli, Gu Hong, and Su Weimin. Angle estimation using ESPRIT without
pairing in MIMO radar. Electronics Letters, 44(24):1422–1423, 2008. (Cited on
page 29.)
[53] Ming Jin, Guisheng Liao, and Jun Li. Joint DOD and DOA estimation for bistatic
MIMO radar. Signal Processing, 89(2):244–251, 2009. (Cited on page 29.)
[54] Jinli Chen, Hong Gu, and Weimin Su. A new method for joint DOD and DOA
estimation in bistatic MIMO radar. Signal Processing, 90(2):714–718, 2010. (Cited
on page 29.)
[55] Xiaofei Zhang, Lingyun Xu, Lei Xu, and Dazhuan Xu. Direction of departure
(DOD) and direction of arrival (DOA) estimation in MIMO radar with reduced-
dimension MUSIC. Communications Letters, IEEE, 14(12):1161–1163, 2010. (Cited
on page 29.)
[56] Mohamed Laid Bencheikh, Yide Wang, and Hongyang He. Polynomial root find-
ing technique for joint DOA DOD estimation in bistatic MIMO radar. Signal Pro-
cessing, 90(9):2723–2730, 2010. (Cited on page 29.)
[57] Mohamed Laid Bencheikh and Yide Wang. Joint DOD-DOA estimation us-
ing combined ESPRIT-MUSIC approach in MIMO radar. Electronics Letters,
46(15):1081–1083, 2010. (Cited on page 29.)
[58] Mohamed Laid Bencheikh and Yide Wang. Combined ESPRIT-rootMUSIC for
DOA-DOD estimation in polarimetric bistatic MIMO radar. Progress In Electro-
magnetics Research Letters, 22:109–117, 2011. (Cited on page 29.)
116 Bibliography
[59] Mohamed Laïd Bencheikh. Exploitation des propriétiés des signaux dans les sys-
tèmes RADAR MIMO pour la détection et la localisation. PhD thesis, Ecole polytech-
nique de l’université de Nantes, Nantes, France, Julliet 2011. (Cited on pages 29
and 59.)
[60] Raymond B Cattell. “Parallel proportional profiles” and other principles for de-
termining the choice of factors by rotation. Psychometrika, 9(4):267–283, 1944.
(Cited on page 29.)
[61] John B Carroll. An analytical solution for approximating simple structure in
factor analysis. Psychometrika, 18(1):23–38, 1953. (Cited on page 29.)
[62] Henry F Kaiser. The varimax criterion for analytic rotation in factor analysis.
Psychometrika, 23(3):187–200, 1958. (Cited on page 29.)
[63] Ledyard R Tucker. Some mathematical notes on three-mode factor analysis. Psy-
chometrika, 31(3):279–311, 1966. (Cited on page 29.)
[64] George M Bergman. Ranks of tensors and change of base field. Journal of Algebra,
11(4):613–621, 1969. (Cited on page 29.)
[65] Richard A Harshman. Foundations of the PARAFAC procedure: models and
conditions for an “explanatory" multimodal factor analysis. In Working Papers in
Phonetics (WPP), 16, pages 1–84. Department of Linguistics, University of Califor-
nia at Los Angeles (UCLA), December 1970. (Cited on pages 29, 30, 85, and 86.)
[66] J Douglas Carroll and Jih-Jie Chang. Analysis of individual differences in multi-
dimensional scaling via an N-way generalization of “Eckart-Young" decomposi-
tion. Psychometrika, 35(3):283–319, 1970. (Cited on pages 29, 85, and 86.)
[67] Richard A Harshman. PARAFAC2: Mathematical and technical notes. UCLA
working papers in phonetics, 22:30–44, 1972. (Cited on page 29.)
[68] Joseph B Kruskal. More factors than subjects, tests and treatments: an indeter-
minacy theorem for canonical decomposition and individual differences scaling.
Psychometrika, 41(3):281–293, 1976. (Cited on page 30.)
[69] Joseph B Kruskal. Three-way arrays: rank and uniqueness of trilinear decom-
positions, with application to arithmetic complexity and statistics. Linear algebra
and its applications, 18(2):95–138, 1977. (Cited on pages 30 and 86.)
[70] Joseph B Kruskal. Rank, decomposition, and uniqueness for 3-way and N-way
arrays. Multiway data analysis, 33:7–18, 1989. (Cited on pages 30 and 86.)
[71] Richard A Harshman and Margaret E Lundy. PARAFAC: Parallel factor analysis.
Computational Statistics & Data Analysis, 18(1):39–72, 1994. (Cited on page 30.)
[72] Rasmus Bro. PARAFAC. tutorial and applications. Chemometrics and intelligent
laboratory systems, 38(2):149–171, 1997. (Cited on page 30.)
Bibliography 117
[73] Rasmus Bro. Multi-way analysis in the food industry: models, algorithms, and applica-
tions. PhD thesis, Universiteit van Amsterdam, Amsterdam, Netherlands, 1998.
(Cited on page 30.)
[74] Henk AL Kiers, Jos MF Ten Berge, and Rasmus Bro. PARAFAC2-Part I. A direct
fitting algorithm for the PARAFAC2 model. Journal of Chemometrics, 13(3-4):275–
294, 1999. (Cited on page 30.)
[75] Rasmus Bro, Claus A Andersson, and Henk AL Kiers. PARAFAC2-Part II. Mod-
eling chromatographic data with retention time shifts. Journal of Chemometrics,
13(3-4):295–309, 1999. (Cited on page 30.)
[76] R Bro, ND Sidiropoulos, and GB Giannakis. Optimal joint azimuth-elevation
and signal-array response estimation using parallel factor analysis. In Conference
Record of the Thirty-Second Asilomar Conference on Signals, Systems & Comput-
ers, volume 2, pages 1594–1598. IEEE, 1998. (Cited on page 30.)
[77] Nicholas D Sidiropoulos, Rasmus Bro, and Georgios B Giannakis. Parallel fac-
tor analysis in sensor array processing. IEEE Transactions on Signal Processing,
48(8):2377–2388, 2000. (Cited on page 30.)
[78] R Bro, N D Sidiropoulos, and G B Giannakis. A fast least squares algorithm for
separating trilinear mixtures. In International Workshop on Independent Component
Analysis and Blind Separation, pages 11–15, 1999. (Cited on pages 30 and 86.)
[79] Claus A Andersson and Rasmus Bro. The N-way toolbox for MATLAB. Chemo-
metrics and Intelligent Laboratory Systems, 52(1):1–4, 2000. (Cited on page 30.)
[80] Henk AL Kiers. Towards a standardized notation and terminology in multiway
analysis. Journal of chemometrics, 14(3):105–122, 2000. (Cited on page 30.)
[81] Nicholas D Sidiropoulos and Rasmus Bro. On the uniqueness of multilinear de-
composition of N-way arrays. Journal of chemometrics, 14(3):229–239, 2000. (Cited
[82] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications.
SIAM review, 51(3):455–500, 2009. (Cited on page 30.)
[83] Pierre Comon. Tensor decompositions, state of the art and applications. arXiv
preprint arXiv: 0905.0454, 2009. (Cited on page 30.)
[84] Gilbert Strang. Introduction to Linear Algebra. Wellesley-Cambridge Press, 2003.
(Cited on page 30.)
[85] Gilbert Strang. Linear algebra and its applications. Belmont, CA: Thomson,
Brooks/Cole, 2006. (Cited on page 30.)
[86] Nicolaas Klaas M Faber, Rasmus Bro, and Philip K Hopke. Recent developments
in CANDECOMP/PARAFAC algorithms: a critical review. Chemometrics and
Intelligent Laboratory Systems, 65(1):119–137, 2003. (Cited on page 30.)
118 Bibliography
[87] Claus A Andersson and Rasmus Bro. Improving the speed of multi-way algo-
rithms: Part I. Tucker3. Chemometrics and intelligent laboratory systems, 42(1):93–
103, 1998. (Cited on page 30.)
[88] Rasmus Bro and Claus A Andersson. Improving the speed of multiway algo-
rithms: Part II. Compression. Chemometrics and Intelligent Laboratory Systems,
42(1):105–113, 1998. (Cited on page 30.)
[89] CN Ho, GD Christian, and ER Davidson. Application of the method of rank an-
nihilation to fluorescent multicomponent mixtures of polynuclear aromatic hy-
drocarbons. Analytical Chemistry, 52(7):1071–1079, 1980. (Cited on page 30.)
[90] Avraham Lorber. Quantifying chemical composition from two-dimensional data
arrays. Analytica chimica acta, 164:293–297, 1984. (Cited on page 30.)
[91] Eugenio Sanchez and Bruce R Kowalski. Generalized rank annihilation factor
analysis. Analytical Chemistry, 58(2):496–499, 1986. (Cited on page 30.)
[92] Eugenio Sanchez and Bruce R Kowalski. Tensorial resolution: a direct trilinear
decomposition. Journal of Chemometrics, 4(1):29–45, 1990. (Cited on page 30.)
[93] Myriam Rajih, Pierre Comon, and Richard A Harshman. Enhanced line search:
A novel method to accelerate PARAFAC. SIAM Journal on Matrix Analysis and
Applications, 30(3):1128–1147, 2008. (Cited on page 30.)
[94] Dimitri Nion and Lieven De Lathauwer. An enhanced line search scheme for
complex-valued tensor decompositions. Application in DS-CDMA. Signal Pro-
cessing, 88(3):749–755, 2008. (Cited on page 30.)
[95] Anh-Huy Phan, Petr Tichavsky, and Andrzej Cichocki. Fast alternating LS al-
gorithms for high order CANDECOMP/PARAFAC tensor factorizations. IEEE
Transactions on Signal Processing, 61(19):4834–4846, 2013. (Cited on page 30.)
[96] Dimitri Nion and Nicholas D Sidiropoulos. A PARAFAC-based technique for
detection and localization of multiple targets in a MIMO radar system. In IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages
2077–2080, Taipei, Taiwan, April 2009. (Cited on pages 31, 85, and 86.)
[97] Dimitri Nion and Nicholas D Sidiropoulos. Tensor algebra and multidimen-
sional harmonic retrieval in signal processing for MIMO radar. IEEE Transac-
tions on Signal Processing, 58(11):5693–5705, November 2010. (Cited on pages 31
and 44.)
[98] Jian-Yun Zhang, Zhi-Dong Zheng, and Xiao-Bo Li. An algorithm for DOD-DOA
and Doppler frequency jointly estimating of bistatic MIMO radar. Dianzi Yu
Xinxi Xuebao(Journal of Electronics and Information Technology), 32(8):1843–1848,
Bibliography 119
[99] Nan Wang, Wenguang Wang, Fan Zhang, and Yunneng Yuan. The PARAFAC-
MUSIC algorithm for DOA estimation with Doppler frequency in a MIMO radar
system. International Journal of Antennas and Propagation, 2014, 2014. (Cited on
page 31.)
[100] Renzheng Cao, Xiaofei Zhang, and Chenghua Wang. Reduced-dimensional
PARAFAC-based algorithm for joint angle and Doppler frequency estimation in
monostatic MIMO radar. Wireless Personal Communications, pages 1–19, 2014.
(Cited on page 31.)
[101] Sebastian Miron, Song Yang, David Brie, and Kainam Wong. A multilinear ap-
proach of direction finding using a sensor-array with multiple scales of spatial
invariance. IEEE Transactions on Aerospace and Electronic Systems, page 00, 2015.
(Cited on page 31.)
[102] Youcef Begriche, Messaoud Thameri, and Karim Abed-Meraim. Exact condi-
tional and unconditional Cramér–Rao bounds for near field localization. Digital
Signal Processing, 31:45–58, 2014. (Cited on pages 31 and 38.)
[103] Parth Raj Singh, Yide Wang, and Pascal Chargé. Bistatic MIMO radar for near
field source localisation using PARAFAC. Electronics Letters, 52(12):1060–1061,
[104] Parth Raj Singh, Yide Wang, and Pascal Chargé. An exact model-based method
for near-field sources localization with bistatic MIMO system. Sensors, 17(4):723,
[105] Parth Raj Singh, Yide Wang, and Pascal Chargé. A correction method for the
near field approximated model based localization techniques. Digital Signal Pro-
cessing, 67:76–80, 2017. (Cited on pages 34 and 105.)
[106] Parth Raj Singh, Yide Wang, and Pascal Chargé. Performance enhancement of
approximated model based near field sources localisation techniques. IET Signal
Processing, 11(7):825–829, 2017. (Cited on pages 34 and 105.)
[107] Parth Raj Singh, Yide Wang, and Pascal Chargé. Near field targets localization
using bistatic MIMO system with symmetric arrays. In 25th European Signal Pro-
cessing Conference (EUSIPCO), pages 2467–2471, Kos island, Greece, August 2017.
EURASIP. (Cited on page 34.)
[108] Parth Raj Singh, Yide Wang, and Pascal Chargé. Near field targets localization
using bistatic MIMO system with spherical wavefront based model. In 25th
European Signal Processing Conference (EUSIPCO), pages 2472–2476, Kos island,
Greece, August 2017. EURASIP. (Cited on page 34.)
[109] Peter Swerling. Probability of detection for fluctuating targets. IRE Transactions
on Information Theory, 6(2):269–308, April 1960. (Cited on page 36.)
120 Bibliography
[110] Eran Fishler, Alexander M. Haimovich, Rick S. Blum, Leonard J. Cimini, Dmitry
Chizhik, and Reinaldo A. Valenzuela. Spatial diversity in radars-models and de-
tection performance. IEEE Transactions on Signal Processing, 54(3):823–838, March
[111] Peter Swerling. Radar probability of detection for some additional fluctuating
target cases. IEEE Transactions on Aerospace and Electronic Systems, 33(2):698–709,
[112] Haidong Yan, Jun Li, and Guisheng Liao. Multitarget identification and localiza-
tion using bistatic MIMO radar systems. EURASIP Journal on Advances in Signal
Processing, 2008(1):1–8, 2007. (Cited on pages 36 and 85.)
[113] Steven M Kay. Fundamentals of Statistical Signal Processing: Estimation Theory,
chapter 3. Cramer-Rao Lower Bound, 15. Extensions for Complex Data and Pa-
rameters, pages 47, 565. PTR Prentice-Hall, Englewood Cliffs, New Jersey, 1993.
(Cited on page 49.)
[114] Petre Stoica, E. G. Larsson, and A. B. Gershman. The stochastic CRB for array
processing: A textbook derivation. Signal Processing Letters, IEEE, 8(5):148–150,
May 2001. (Cited on page 49.)
[115] José Tribolet. A new phase unwrapping algorithm. IEEE Transactions on Acoustics
Speech and Signal Processing, 25(2):170–177, April 1977. (Cited on pages 54, 83,
and 87.)
[116] Feifei Gao and Alex B Gershman. A generalized ESPRIT approach to direction-
of-arrival estimation. IEEE Signal Processing Letters, 12(3):254–257, March 2005.
(Cited on page 59.)
[117] Donald Shepard. A two-dimensional interpolation function for irregularly-
spaced data. In 23rd ACM national conference, pages 517–524, New York, USA,
August 1968. ACM. (Cited on page 68.)
[118] Gene H Golub and Charles Van Loan. Total least squares. In Smoothing Techniques
for Curve Estimation, pages 69–76. Springer, 1979. (Cited on page 73.)
[119] Yariv Ephraim and Harry L. Van Trees. A signal subspace approach for speech
enhancement. IEEE Transactions on Speech and Audio Processing, 3(4):251–266, July
[120] Andrzej Cichocki, Danilo P. Mandic, Anh Huy Phan, Cesar F. Caiafa, Guoxu
Zhou, Qibin Zhao, and Lieven De Lathauwer. Tensor decompositions for signal
processing applications: From two-way to multiway component analysis. IEEE
Signal Processing Magazine, 32(2):145–163, 2015. (Cited on page 85.)
Thèse de Doctorat
Parth Raj S INGH

Localisation de Source par les Systèmes MIMO
Source Localization with MIMO Systems
Résumé Abstract
Dans cette thèse, nous considérons la dernière géné- Sources localization is used in radar, sonar, and
ration du radar. Il s’agit d’un radar MIMO bistatique telecommunication. Radar has numerous civilian and
qui est composé de plusieurs antennes d’émission et military applications. Radar system has gone through
de réception. Pour ce système, les antennes émet- many developments over the last few decades and
trices transmettent des signaux linéairement indépen- reached the latest version known as MIMO radar. A
dants afin qu’ils puissent être identifiés à l’aide d’un MIMO radar is composed of multiple transmitting and
banc de filtres adaptés au niveau des antennes de receiving antennas like a conventional phased array
réception. Les signaux filtrés sont alors traités pour radar. However, its transmitting antennas transmit lin-
extraire les paramètres des cibles, tels que les DOA, early independent signals so that they can be easily
DOD, vitesse, etc. Un radar MIMO bistatique offre une identified by the matched filters bank at its receiving
grande diversité spatiale et une excellente identifiabi- end. The matched filtered signals are then processed
lité des paramètres, etc., ce qui nous a incités à l’uti- to extract the ranges, DOAs, DODs, velocities, etc. of
liser dans ce travail. La situation en champ lointain the targets. A bistatic MIMO radar system provides
d’un radar MIMO bistatique est largement traitée dans high resolution, spatial diversity, parameter identifiabil-
la littérature. Mais, peu de travaux existe sur la situa- ity, etc. which inspired us to use it in this work. There
tion en champ proche, c’est ce qui a motivé le tra- are many existing methods to deal with the far field re-
vail de cette thèse. La localisation de cibles en champ gion of MIMO radar system. However, little work can
proche est importante en raison de nombreuses ap- be found on the near field region of a bistatic MIMO
plications à l’intérieur des constructions. A ce sujet, la radar which motivated the work in this thesis. Near
plupart des méthodes actuelles utilisent l’approxima- field targets localization is also important because of
tion de Fresnel dans laquelle le front d’onde sphérique many indoor applications. Most of the existing near
des signaux reçus est supposé quadrique plutôt que field sources localization techniques use Fresnel ap-
planaire comme en champ lointain. Dans ce travail de proximation in which the real spherical wavefront is as-
thèse, nous avons proposé une nouvelle méthode de sumed quadric unlike planar in far field situation. In
localisation des cibles en champ proche qui utilise l’ap- this work we have proposed a novel near field targets
proximation de Fresnel. Celle-ci conduit à une estima- localization method using Fresnel approximation. The
tion biaisée des paramètres de localisation car en réa- Fresnel approximation leads to a biased estimation of
lité le front d’onde est sphérique. Nous avons proposé the location parameters because the true wavefront is
alors deux méthodes de correction pour réduire les ef- spherical. Consequently, we have proposed two cor-
fets de l’approximation de Fresnel et deux autres mé- rection methods to reduce the effects of Fresnel ap-
thodes qui utilisent directement le modèle exacte basé proximation and other two methods which directly use
sur le front d’onde sphérique. the exact signal model based on spherical wavefront.
Mots clés Key Words

Système MIMO bistatique, ULA, champ proche, Bistatic MIMO system, ULA, near field, Fresnel
approximation de Fresnel, distance, DOA, DOD. approximation, range, DOA, DOD.
L’UNIVERSITÉ NANTES ANGERS LE MANS

Thèse de Doctorat: Parth Raj S

Uploaded by

Copyright:

Available Formats

Thèse de Doctorat: Parth Raj S

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thèse de Doctorat: Parth Raj S

Uploaded by

Copyright:

Available Formats

Thèse de Doctorat

Parth Raj S INGH

École doctorale : Sciences et technologies de l’information, et mathématiques (STIM-503)

Soutenue le 12 October 2017

Source Localization with MIMO Systems

During this three-year research work at Institute of Electronics and Telecommunica-

réseau d’antennes d’émission (DOD) ; la direction d’arrivée par rapport au réseau

• La première méthode de correction utilise deux LUTs correspondant aux réseaux

méthodes permettant d’obtenir les paramètres de localisation biaisés. Chaque

estimer les paramètres de localisation.

• La deuxième méthode basée sur le modèle exacte utilise la décomposition des

1.5 Cramér-Rao Lower Bound (CRLB) . . . . . . . . . . . . . . . . . . . . . . 49

2 Approximated Model Based Methods 51

4 Accurate Model Based Methods 81

A Biased Location Parameters 95

A.1.6 Residue Vector Approximation . . . . . . . . . . . . . . . . . . . . 100

B Correction Methods for Single Array 105

1 Passive and active radar systems schematic. . . . . . . . . . . . . . . . . 25

1.1 Transmitting ULA parameters . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.1 Variation of unambiguous region of θe p with de . It is same for θr p with

3.1 Discretized Fresnel region of an array. . . . . . . . . . . . . . . . . . . . . 67

B.1 ξer p of [2] in noiseless condition; d = λ/4, L = 103 , and N̆ = 4. . . . . . . 106

2.1 Proposed Approximated Model Based Method. . . . . . . . . . . . . . . . 61

CANDECOMP CANonical DECOMPosition

DOA Direction Of Arrival

GRAM Generalized Rank Annihilation Method

MATLAB MATrix LABoratory

PARAFAC PARAllel FACtor

RADAR RAdio Detection And Ranging

RCS Radar Cross Section

SNR Signal-to-Noise Ratio

TALS Trilinear Alternating Least Squares

ULA Uniform Linear Array

diag{ a} Diagonal matrix with the elements of a as its diagonal elements

played a significantly important role in the development of radar [4, 5].

1.2 Passive and Active Radar Systems

Figure 1 – Passive and active radar systems schematic.

1.3 Array Based Radar

Because of the influence of the military applications on the initial development of

2.2 Classification of MIMO Radar Systems

MIMO radar Bistatic

Figure 2 – Types of MIMO radar.

3 State of the Art

use a two-dimensional estimator or use two one-dimensional estimators. Maximum

4 Organization of this Dissertation

the two methods in the presence of two targets.

As mentioned before, the Fresnel approximation introduces biases in the estimated

Finally, we conclude the work with perspectives and future works.

5 List of Related Publications

1.1 Target Model

1.2 Modeling of Bistatic MIMO System

y(t) = A s(t) + w(t) (1.1)

with A = Ae Ar where Ae ∈ C M× P and Ar ∈ C N × P contain the directional vectors of

Figure 1.1 – Transmitting ULA parameters

1.3 Modeling of Near Field Targets

1.3.1 Fresnel Region

1.3.2 Wavefront Approximation in Near Field

Figure 1.2 – Fresnel region.

Equation (1.5) can be rewritten as

(n − no )2 d2r + ρ2r p (n − no )2 d2r − ρ2r p

(n − no )2 d2r + ρ2r p (n − no )2 d2r − ρ2r p