ANALYZING AND MODELING LOW-COST MEMS IMUS FOR USE IN AN

INERTIAL NAVIGATION SYSTEM

by

Justin Michael Barrett

A Thesis

Submitted to the Faculty

of the

WORCESTER POLYTECHNIC INSTITUTE

in partial fulfillment of the requirements for the

Degree of Master of Science

in

Robotics Engineering

May 2014

APPROVED:

_______________________________________
Professor Michael A. Gennert, Major Advisor, Head of Program

_______________________________________
Professor William R. Michalson

_______________________________________
Dr. Julian L. Center, Jr.
 Abstract

Inertial navigation is a relative navigation technique commonly used by autonomous vehicles to

determine their linear velocity, position and orientation in three-dimensional space. The basic premise of
inertial navigation is that measurements of acceleration and angular velocity from an inertial
measurement unit (IMU) are integrated over time to produce estimates of linear velocity, position and
orientation. However, this process is a particularly involved one. The raw inertial data must first be
properly analyzed and modeled in order to ensure that any inertial navigation system (INS) that uses the
inertial data will produce accurate results.
This thesis describes the process of analyzing and modeling raw IMU data, as well as how to use
the results of that analysis to design an INS. Two separate INS units are designed using two different
micro-electro-mechanical system (MEMS) IMUs. To test the effectiveness of each INS, each IMU is
rigidly mounted to an unmanned ground vehicle (UGV) and the vehicle is driven through a known test
course. The linear velocity, position and orientation estimates produced by each INS are then compared to
the true linear velocity, position and orientation of the UGV over time. Final results from these
experiments include quantifications of how well each INS was able to estimate the true linear velocity,
position and orientation of the UGV in several different navigation scenarios as well as a direct
comparison of the performances of the two separate INS units.

ii
 Acknowledgements

First and foremost I would like to thank my family and friends for their unwavering support over
the last eight years of my academic career at WPI and the many years preceding it. You have always been
confident in my abilities even when I wasn’t, and without you this thesis simply would not have been
possible.

Second, I would like to thank my research advisor Professor Michael A. Gennert and the other
two members of my thesis committee Professor William R. Michalson and Dr. Julian L. Center, Jr. for
consistently keeping me and my research on the right track. Thank you for constantly answering all of my
questions and for providing me with praise when I earned it and constructive criticism when I needed it.

Third, I would like to thank Autonomous Exploration, Inc. and NASA for providing the funding
for this research. Without their financial contributions I would have missed out on this valuable research
experience that has taught me so much. I would also like to thank Bryan W. Welch of the NASA Glenn
Research Center for being our team’s NASA contact and for periodically providing us with advice on
how to troubleshoot our technical problems.

Finally, I would like to thank my fellow team members James F. Kirk and Michael D. Audi for
their contributions to the larger research project that this thesis is a part of. Your contributions to the
project made my life significantly easier, and working with you was always an enjoyable and worthwhile
experience.

iii
Table of Contents

List of Figures ............................................................................................................................... vi

List of Tables ................................................................................................................................ viii
Section 1 – Introduction ................................................................................................................ 1
Section 2 – Background Research ................................................................................................ 6
Section 3 – IMU Errors and Error Models .................................................................................... 10
Section 3.1 – Deterministic Errors ................................................................................... 10
Section 3.1.1 – Static Biases ............................................................................... 11
Section 3.1.2 – Scale Factors and Misalignment Errors ..................................... 12
Section 3.2 – Stochastic Errors ........................................................................................ 15
Section 3.2.1 – Measurement Noise ................................................................... 15
Section 3.2.2 – Turn-On to Turn-On Bias Variation .......................................... 17
Section 3.2.3 – Drifting Biases ........................................................................... 17
Section 4 – Laboratory Experiments............................................................................................. 19
Section 4.1 – 12-Hour Static IMU Data Experiment ....................................................... 19
Section 4.1.1 – Allan Variance Analysis ............................................................ 21
Section 4.1.2 – Power Spectral Density (PSD) Analysis .................................... 25
Section 4.1.3 – Probability Density Function (PDF) Analysis ........................... 28
Section 4.1.4 – Drifting Bias Analysis................................................................ 32
Section 4.2 – Power Cycle Experiment ........................................................................... 34
Section 4.3 – Multi-Angle Experiment ............................................................................ 36
Section 4.3.1 – Linear Least Squares Estimation Analysis................................. 39
Section 4.4 – Rate Table Experiment .............................................................................. 42
Section 4.4.1 – Linear Least Squares Estimation Analysis................................. 46
Section 5 – The Inertial Navigation Algorithm............................................................................. 48
Section 5.1 – Coordinate Frames ..................................................................................... 48
Section 5.2 – Assumptions............................................................................................... 50
Section 5.3 – Inertial Navigation Algorithm Overview ................................................... 51
Section 5.4 – The Initialization Phase.............................................................................. 54
Section 5.5 – The Estimation Phase ................................................................................. 54
Section 5.6 – The Correction Phase ................................................................................. 57
Section 6 – UGV Navigation Test Plan ........................................................................................ 63

iv
 Section 6.1 – The Test Course ......................................................................................... 63
Section 6.2 – Configuring and Driving the UGV ............................................................ 65
Section 7 – Results and Discussion............................................................................................... 68
Section 7.1 – Inertial Data Only ...................................................................................... 68
Section 7.2 – Inertial Data and ZUPTs ............................................................................ 69
Section 7.3 – Inertial Data, ZUPTs and Position and Orientation Fixes.......................... 70
Section 7.4 – SparkFun and ADI IMU Comparison ........................................................ 71
Section 8 – Future Work ............................................................................................................... 74
Section 9 – Conclusion ................................................................................................................. 75
Appendix ....................................................................................................................................... 76
Section A.1 – Full Results of UGV Navigation Testing (Trial #1) ................................. 76
Section A.2 – Full Results of UGV Navigation Testing (Trial #2) ................................. 80
Section A.3 – Full Results of UGV Navigation Testing (Trial #3) ................................. 83
References ..................................................................................................................................... 86

v
List of Figures

Figure 1.1 – A block diagram illustrating the basic operation of a typical INS ............................ 2
Figure 1.2 – The ADI IMU and the SparkFun IMU ..................................................................... 4
Figure 3.1 – An example of an unbiased sine wave and a biased sine wave ................................ 11
Figure 3.2 – An example of a non-scaled sine wave and a scaled sine wave ............................... 12
Figure 3.3 – An example of an IMU axis misalignment ............................................................... 13
Figure 3.4 – An example of a sine wave with and without noise ................................................. 15
Figure 3.5 – An example of a drifting bias in a sample of IMU gyroscope data .......................... 18
Figure 4.1 – The test fixture for the 12-Hour Static IMU Data Experiment ................................. 20
Figure 4.2 – 12-Hour Static IMU Data Experiment test fixture enclosed in incubator ................ 21
Figure 4.3 – The Allan Deviation function of a sample of WGN generated by MATLAB.......... 23
Figure 4.4 – Allan Deviation functions of static data from the ADI IMU .................................... 23
Figure 4.5 – Allan Deviation functions of static data from the SparkFun IMU ........................... 24
Figure 4.6 – The division of a time series into fixed-length overlapping segments ..................... 25
Figure 4.7 – The Welch PSD of a sample of WGN generated by MATLAB............................... 26
Figure 4.8 – Welch PSDs of static data from the ADI IMU ......................................................... 27
Figure 4.9 – Welch PSDs of static data from the SparkFun IMU................................................. 27
Figure 4.10 – Histogram of WGN with a Gaussian distribution fitted to it .................................. 29
Figure 4.11 – Histogram of the measurement noise processes from the ADI IMU ...................... 29
Figure 4.12 – Histogram of the measurement noise processes from the SparkFun IMU ............. 30
Figure 4.13 – ADI IMU histograms with Gaussian distributions fit to them................................ 30
Figure 4.14 – SparkFun IMU histograms with Gaussian distributions fit to them ....................... 31
Figure 4.15 – The test fixture for the Power Cycle Experiment ................................................... 34
Figure 4.16 – Power Cycle Experiment test fixture in a hybridization incubator ......................... 35
Figure 4.17 – Multi-Angle Experiment test fixture with the IMU mounted ................................. 37
Figure 4.18 – Multi-Angle Experiment test fixture in the hybridization incubator ...................... 37
Figure 4.19 – Rate Table Experiment test fixture with the IMU mounted ................................... 43
Figure 4.20 – The IMU data logging peripherals.......................................................................... 44
Figure 4.21 – Rate Table Experiment test fixture enclosed in the hybridization incubator .......... 45
Figure 5.1 – The four coordinate frames used by the inertial navigation algorithm ..................... 48
Figure 5.2 – The positive angular velocity conventions defined by the Body frame ................... 49
Figure 5.3 – A high-level block diagram of the inertial navigation algorithm ............................. 52

vi
Figure 5.4 – Block diagram of the estimation phase of the inertial navigation algorithm ............ 55
Figure 5.5 – Block diagram of the correction phase of the inertial navigation algorithm ............ 58
Figure 6.1 – Position markers outlining the locations of ZUPTs .................................................. 64
Figure 6.2 – The Husky A100 UGV ............................................................................................. 65
Figure 6.3 – The wheels of the Husky A100 UGV coated in duct tape ........................................ 66
Figure 6.4 – SparkFun and ADI IMUs rigidly mounted to the Husky A100 UGV ...................... 66
Figure 6.5 – The Husky A100 UGV positioned at the beginning of the test course..................... 67
Figure 7.1 – Estimated position and orientation (Inertial data only) ............................................ 68
Figure 7.2 – Estimated position and orientation (Inertial data and ZUPTs only) ......................... 69
Figure 7.3 – Estimated position and orientation (All information) ............................................... 70
Figure 7.4 – Comparison of position and orientation estimates.................................................... 71
Figure 7.5 – Comparison of linear velocity estimates................................................................... 72
Figure 7.6 – Errors in linear velocity, position and orientation (ADI).......................................... 72
Figure 7.7 – Errors in linear velocity, position and orientation (SparkFun) ................................. 73
Figure A.1 – Estimated linear velocity of the vehicle (ADI Trial #1) .......................................... 76
Figure A.2 – Estimated linear velocity of the vehicle (SparkFun Trial #1) .................................. 77
Figure A.3 – Estimated position and orientation of the vehicle (ADI Trial #1) ........................... 77
Figure A.4 – Estimated position and orientation of the vehicle (SparkFun Trial #1) ................... 78
Figure A.5 – Linear velocity, position and orientation estimation errors (ADI Trial #1) ............. 78
Figure A.6 – Linear velocity, position and orientation estimation errors (SparkFun Trial #1) .... 79
Figure A.7 – Estimated linear velocity of the vehicle (ADI Trial #2) .......................................... 80
Figure A.8 – Estimated linear velocity of the vehicle (SparkFun Trial #2) .................................. 80
Figure A.9 – Estimated position and orientation of the vehicle (ADI Trial #2) ........................... 81
Figure A.10 – Estimated position and orientation of the vehicle (SparkFun Trial #2) ................. 81
Figure A.11 – Linear velocity, position and orientation estimation errors (ADI Trial #2) ........... 82
Figure A.12 – Linear velocity, position and orientation estimation errors (SparkFun Trial #2) .. 82
Figure A.13 – Estimated linear velocity of the vehicle (ADI Trial #3) ........................................ 83
Figure A.14 – Estimated linear velocity of the vehicle (SparkFun Trial #3) ................................ 83
Figure A.15 – Estimated position and orientation of the vehicle (ADI Trial #3) ......................... 84
Figure A.16 – Estimated position and orientation of the vehicle (SparkFun Trial #3) ................. 84
Figure A.17 – Linear velocity, position and orientation estimation errors (ADI Trial #3) ........... 85
Figure A.18 – Linear velocity, position and orientation estimation errors (SparkFun Trial #3) .. 85

vii
List of Tables

Table 1.1 – Selected performance specifications for the SparkFun and ADI IMUs ..................... 4
Table 4.1 – Allan Deviation function slope values for the ADI and SparkFun IMUs .................. 24
Table 4.2 – PSD slope values for the ADI and SparkFun IMUs .................................................. 28
Table 4.3 – Sample variances of the IMU measurement noise processes..................................... 31
Table 4.4 – Estimated values of drifting bias first-order Markov process parameters ................. 33
Table 4.5 – Turn-on to turn-on bias variation for the ADI and SparkFun IMUs .......................... 36
Table 4.6 – Static biases, scale factors and misalignment errors for the accelerometers.............. 41
Table 4.7 – Static biases, scale factors and misalignment errors for the gyroscopes.................... 47

viii
Section 1 – Introduction

As robotics technology continues to become more prevalent in modern society, the development
and production of autonomous vehicles continues to expand to meet the increasing demand for vehicles
that can safely explore treacherous, inhospitable or hard to reach locations. One of the most crucial
components of any autonomous vehicle, especially one tasked with exploring and/or mapping unknown
environments, is its navigation system. Without the ability to accurately determine its linear velocity,
position and orientation with respect to some larger frame of reference, an autonomous vehicle would be
drastically limited in terms of the tasks it could accurately perform. Therefore, the implementation of an
accurate navigation system on any autonomous vehicle is an important aspect of the system design and is
imperative to the overall ability of the vehicle to function properly.
One type of navigation system commonly implemented on autonomous vehicles is called an
inertial navigation system (INS). Inertial navigation systems measure the accelerations and angular
velocities that an autonomous vehicle experiences (typically in three dimensions) and integrate those
measurements over time to produce estimates of the linear velocity, position and orientation of the
vehicle. Inertial navigation is particularly useful because it can be implemented on almost any type of
vehicle and in almost any type of environment. In comparison, many absolute navigation sensors such as
GPS receivers, Digital Magnetic Compasses (DMCs), etc. have a much more limited range of usefulness
and are typically most effective when used in relatively spacious outdoor environments.
The measurements of acceleration and angular velocity that are required for inertial navigation
are produced by a sensor package called an inertial measurement unit (IMU). There are several different
varieties of IMUs, but one of the more common varieties is the micro-electro-mechanical-system
(MEMS) IMU. MEMS IMUs are very inexpensive to manufacture, which makes them a popular choice
among robotics hobbyists. However, as is the case with any type of sensor, IMUs can never produce
perfect measurements of acceleration and angular velocity. Since the measured accelerations and angular
velocities returned by the IMU are imperfect, it follows that any estimates of linear velocity, position and
orientation based on those measurements will be imperfect as well. The end result is that after a short
period of time, the estimates of linear velocity, position and orientation produced by an INS very quickly
become inaccurate and unusable.
To compensate for the imperfections in the IMU measurements, an INS typically has one or more
secondary navigation sensors that provide direct measurements of the linear velocity, position and/or
orientation of a vehicle. These secondary navigation sensors could be anything from stereo vision
systems, to GPS receivers, to digital magnetic compasses (DMCs) or any other type of sensor that could

1
be used to measure linear velocity, position or orientation. The information from these secondary
navigation sensors is incorporated into the INS using an Extended Kalman Filter (EKF). The EKF
produces correction terms that are used to adjust the initial estimates of linear velocity, position and
orientation calculated from the imperfect IMU measurements. Adding secondary navigation sensors into
an INS greatly increases its ability to produce accurate estimates of the linear velocity, position and
orientation of a vehicle over long periods of time. The block diagram in Figure 1.1 illustrates the basic
operation of a typical INS.

Figure 1.1: A block diagram illustrating the basic operation of a typical INS.

2
 The goal of this thesis is to design, implement and test two separate INS units based around two
low-cost MEMS IMUs. The results of the testing procedures are then used to determine how well each
individual INS performs as well as to determine which IMU is better suited for implementation in an INS.
This goal is achieved by following these basic steps:

• Each IMU is analyzed and modeled using a series of laboratory experiments designed to isolate
and characterize the behavior of several well-documented types of errors common to MEMS
IMUs.
• An inertial navigation algorithm is designed that incorporates the information and error models
produced by the laboratory experiments.
• To collect test data for the inertial navigation algorithm, the two IMUs are rigidly mounted to an
unmanned ground vehicle (UGV) and are configured to log inertial data at 25 Hz while the UGV
is driven via remote control through a predefined test course. The test course is designed such that
the true linear velocity, position and orientation of the UGV can be accurately measured at
periodic intervals along the course. This truth data will serve as the secondary navigation
information that gets fed into the EKF component of the inertial navigation algorithm.
• The inertial information recorded from each IMU during the navigation tests is then post
processed using the inertial navigation algorithm. The resulting navigation information is
analyzed and compared to the true linear velocity, position and orientation of the UGV to
determine which INS was able to more accurately estimate the linear velocity, position and
orientation of the UGV over the course of the navigation tests.

The procedures described above are implemented using two low-cost MEMS IMUs. The first is a
hobbyist-grade IMU from the website SparkFun.com, and the second is a commercial-grade IMU from
the company Analog Devices, Inc. (ADI). The two IMUs can be seen in Figure 1.2. Both the SparkFun
and ADI IMUs contain a three-axis accelerometer, a three-axis gyroscope and an embedded temperature
sensor that is capable of measuring the internal temperature of the IMU package in degrees Celsius. Some
performance specifications from the SparkFun and ADI IMU datasheets are listed in Table 1.1.
As Table 1.1 illustrates, the specifications for the ADI IMU are superior to the specifications for
the SparkFun IMU in some categories, but not all of them. Given that the ADI IMU costs four times as
much as the SparkFun IMU, it is intuitive to think that the navigation information produced using the
ADI IMU will be much better than the navigation information produced using the SparkFun IMU.
However, the relative quality of two IMUs cannot be determined just by looking at the differences in their

3
 Figure 1.2: The ADI IMU (on the left) and the SparkFun IMU (on the right).

SparkFun IMU ADI IMU

Unit Information
Model # SEN-10736 ADIS16364
Cost Per Unit $125 $500
Accelerometers
Range ±2g ±5g
Resolution 3.9 mg/LSB 1 mg/LSB
± 40 mg (X and Y axis) ± 80 mg (Z
0g Bias Level 8 mg
axis)
Noise Performance < 3.9 mg RMS 5 mg RMS
Gyroscopes
Range ± 2000 °/sec ± 350 °/sec
Resolution 0.0696 °/sec/LSB 0.05 °/sec/LSB
0g Bias Level ± 40 °/sec ± 3°/sec
Noise Performance 0.38 °/sec RMS 0.8 °/sec RMS

Table 1.1: Selected performance specifications for the SparkFun and ADI IMUs [1-3].

price tags. Only after performing the experiments outlined in this thesis can it be definitively said that one
IMU outperforms the other in terms of the ability to provide accurate navigation information.

4
 The remaining sections of this thesis are briefly summarized in the bulleted list below:

• Section 2 presents the background research performed for this thesis. Each of the five papers
presented represents some previous research that has been conducted in the field of IMU error
modeling, IMU error calibration or inertial navigation.
• Section 3 discusses the two main categories of IMU errors (deterministic and stochastic) in detail
and breaks both categories down into several subcategories. The mathematical representations of
each type of IMU error are also discussed.
• Section 4 explains the laboratory experiments designed to isolate and quantify each of the IMU
errors discussed in Section 3. Each experiment is described in detail, including its purpose, how it
is set up and the end results that are produced.
• Section 5 describes the inertial navigation algorithm in more detail. A brief explanation of the
basic principles of inertial navigation is provided, the assumptions made by the algorithm are
stated and each individual component of the inertial navigation algorithm is thoroughly
explained.
• Section 6 outlines the INS testing process. The predefined test course for the UGV is presented
along with a description of how the secondary navigation information for the EKF component of
the inertial navigation algorithm is recorded.
• Section 7 presents and discusses the results obtained from the UGV navigation testing. The
results are presented visually, displaying the estimated position and orientation of the UGV and
the true position and orientation of the UGV on the same graph. A comparison of the two INS
units is also presented to determine which INS was able to more accurately estimate the linear
velocity, position and orientation of the UGV over the course of the navigation testing.
• Section 8 lists some potential ideas for future work, and Section 9 restates the goal of this thesis
and summarizes the general conclusions that can be made about the results that were obtained.
• The Appendix contains supplemental figures detailing the complete results from all of the UGV
navigation testing trials.

5
Section 2 – Background Research

The field of inertial navigation has been extensively researched and well documented over the
course of the last fifty years. There are many published papers that document a variety of different ways
to model the various types of errors that commonly appear in inertial data. This section presents five of
these papers and explains how they are related to the work demonstrated in this thesis.
The first paper is titled “Calibration of Accelerometer Triad on an IMU with Drifting Z-
Accelerometer Bias” [4]. The goal of this paper was to design a test that would quantify a series of error
parameters (bias, scale factor, dis-symmetry and non-orthogonality errors) for a three-axis accelerometer
with the prior knowledge that the Z-axis accelerometer bias varies with time. Normally, this type of test
would be conducted under the assumption that all of the error parameters being estimated remain constant
during the duration of the test. However, because the Z-axis accelerometer bias is known to vary with
time, this assumption cannot be made. A method for estimating the Z-axis accelerometer bias is proposed
that takes advantage of the correlation between the Z-axis accelerometer bias and the non-orthogonality
errors of all three accelerometer axes. If the non-orthogonality errors are assumed to be constant and
negligibly close to zero, then any non-zero values that appear in the estimated non-orthogonality terms
must be a result of the drift in the Z-axis accelerometer bias. This relationship between the non-
orthogonality errors and the Z-axis accelerometer bias drift can be used to estimate the steady-state value
of the Z-axis accelerometer bias. The paper proves the validity of this concept through the use of a
computer simulation.
This paper proved to be a useful resource because it describes a multi-angle experiment that is
very similar to the one described in Section 4.3. The assertion that the Z-axis accelerometer bias drift can
cause inaccuracies in all of the other error parameter estimates seems difficult to believe, although that
assertion may be based on how long the IMU is left in each orientation. The Multi-Angle Experiment
conducted in this thesis records five minutes worth of data at each orientation, so the amount of bias drift
that occurs at any given orientation is most likely small enough to be considered negligible. If the bias
drift is assumed to be negligible, then the experiment becomes much simpler because it can be assumed
that all of the error parameters remain constant over the duration of the test.
The second paper is titled “Research on Auto Compensation Technique of Strap-Down Inertial
Navigation Systems” [5]. The goal of this paper was to show that the heading estimates produced by an
INS could be improved by mounting it to a vehicle using a platform that constantly rotates about the
gravity vector. This is opposed to the more conventional method of mounting an INS rigidly to the body
of a vehicle. As the platform rotates, it generates a constant angular velocity. This constant angular

6
velocity is measured by the INS and appears in the data as a change in one of the gyroscope biases. This
additional bias term from the rotating platform ensures that even when the heading of the vehicle is
constant, there is always some non-zero component of angular velocity that the INS can measure. Also,
because the angular velocity of the rotating platform is known, error calibration routines can be run on the
INS at all times. This results in more accurate heading estimates over longer periods of time. The paper
goes on to prove this concept through the use of a computer simulation that models the implementation of
this type of INS mounting strategy on a vehicle travelling with a fixed heading over a twenty four hour
time period.
The rotating INS platform described in this paper is very similar to the Rate Table Experiment
described in Section 4.4. Although the method described in the paper is for a continuous calibration
process, the core idea of exposing the gyroscopes to a constant angular velocity in order to determine
scale factor and misalignment errors can be adapted for use in a more comprehensive laboratory
experiment. This paper provides some useful information on how to structure the Rate Table Experiment,
and how to interpret the results that it produces.
The third paper is titled “Indoor 3D Position Estimation Using Low-Cost Inertial Sensors and
Marker-Based Video-Tracking” [6]. The goal of this paper was to develop an indoor object-tracking
system using a low-cost IMU and an external outside-looking-in video tracking system. The paper states
that an IMU can be used to provide a rough estimate of the position, velocity, and orientation of an object,
but that it is also subject to errors caused by biases and bias drifts. In order to compensate for these errors,
an external video tracking system is used to provide corrections for the IMU estimates. The data returned
from the IMU and the video tracking system is synchronized and fused together using an EKF, which
results in a more accurate estimate of the detected object’s position, velocity, and orientation. The system
was first tested on a set of experimental data to provide a proof of concept test for the system design, and
then was subsequently tested on an actual object using the system hardware to provide proof that the
system was capable of tracking real world objects. For the real world test, a test subject’s hand was used
as the object to be tracked by the system. The IMU was strapped to the subject’s wrist, and a fiducial
marker cube was attached to the subject’s arm to act as a target for the video tracking system to track. The
results of both the simulated and real world experiments showed that the system was capable of
accurately estimating the three-dimensional position and orientation of a detected object, and the position
estimates were accurate to within a few centimeters of the true position values. The results also showed
that the combination of the IMU and video tracking sensors resulted in a more robust and accurate system
than if either of the two sensors were used on their own.
The primary reason for including this paper is because the structure of the EKF-based navigation
system is very similar to the inertial navigation algorithm that is presented in this thesis. And while the

7
paper didn’t mention any other IMU calibration techniques other than the bias tracking and compensation
mechanism of the EKF, it is still a valuable resource to have because the navigation system that it
presents is similar in design to the inertial navigation algorithm that this thesis is based around.
The fourth paper is titled “Thermal Calibration of MEMS Inertial Sensors for an FPGA-Based
Navigation System” [7]. This paper focused on both the development and calibration of a custom IMU
that uses an FPGA-based processor to produce real-time navigation information. It briefly discussed the
process of developing the IMU which contained low-cost accelerometers and gyroscopes from Analog
Devices, Inc., and the FPGA-based processor which was used for both data collection and navigation
calculations. However, the majority of the paper discussed the calibration routines that are required to
compensate for the deterministic and stochastic errors in the IMU data. Error sources such as biases, bias
drifts, scale factors, misalignment errors, and thermal sensitivity were all discussed, and methods for
compensating for them were explained. Because thermal sensitivity affects all of the other error
parameters, all of the IMU calibration testing was performed under the umbrella of temperature regulated
testing. The paper concluded by presenting the results of the calibration experiments, and by showing
how accurately they were able to compensate for the IMU errors.
This paper contains useful information about IMU calibration in general (Allan Variance
analysis, multi-angle experiments, etc.), but the primary reason for including it is because of the focus that
it puts on thermal sensitivity and how it can affect IMU calibration. Because MEMS sensors are so
sensitive to changes in temperature, it is important to realize that any calibration routines performed on an
IMU will only yield results that will be usable at the temperature that the test was performed at.
Therefore, in order to accurately compensate for IMU errors at different temperatures, calibration routines
must be performed at a variety of different temperatures so that a model can be constructed that
accurately describes how each of the different IMU errors changes with respect to temperature. This paper
is a valuable reference to have because of the information that it provides regarding the effects of
temperature variation on the behavior of IMU errors.
The fifth and final paper is titled “Estimation of Deterministic and Stochastic IMU Error
Parameters” [8]. This paper breaks down the problems of IMU calibration into two main components:
deterministic errors and stochastic errors. Experiments designed to quantify deterministic errors include
multi-angle and rate table experiments, which expose the IMU to a series of controlled accelerations and
angular velocities in order to determine how much the IMU measurements differ from the true
accelerations and angular velocities that the IMU is actually experiencing. Experiments designed to model
stochastic errors include estimation techniques such as Kalman Filtering, which produces estimates of the
states of a dynamic system by using a model of the system and imperfect measurements of the system
states. Stochastic errors are typically more difficult to compensate for due to their inherently random

8
nature. The paper uses both simulated and real IMU data to test the proposed calibration methods, and
presents the successful results that the researchers were able to obtain. Overall, the results of this paper
showed that it is possible to improve the performance of an IMU by quantifying and modeling the various
deterministic and stochastic errors in the IMU data and then using that information to help compensate for
the negative effects of the errors.
The primary reason for including this paper is that it is very similar to this thesis in terms of what
it sets out to achieve and the methodology that it uses to achieve it. It attempts to estimate a very similar
group of IMU error parameters using a very similar series of laboratory experiments, and it also utilizes
many of the same mathematical concepts (primarily least-squares estimation and Kalman Filtering) that
are used for performing error analyses in this thesis. The only major topic from this thesis that the paper
doesn’t cover is temperature sensitivity testing. Overall, this paper is a valuable reference to have because
it describes several common types of IMU errors and how they behave, and it also provides a lot of
insight into how to design experiments that will both isolate and quantify those errors.

9
Section 3 – IMU Errors and Error Models

There are some IMUs that are capable of producing incredibly accurate measurements of
acceleration and angular velocity. However, there is no IMU that can produce perfect measurements of
acceleration and angular velocity. No matter how good an IMU is, there will always be imperfections in
the inertial data that it returns. These imperfections are caused by several different kinds of errors, and
each has a different effect on the data. IMU errors can be broken up into two primary categories:
deterministic errors and stochastic errors. Each of these categories can then be further broken down into a
series of subcategories, one for each different kind of error. By understanding each of these errors and
modeling their behavior mathematically, it becomes possible to compensate for them and improve the
overall quality of the IMU data.
It should be noted that there are many different types of IMU errors, and only a subset of them are
discussed in this thesis. A substantial amount of time can be devoted to precisely modeling and
understanding the behavior of all of the different types of IMU errors that have been previously
documented. However, after a certain extent the benefit from doing so becomes less and less. Because of
this, a specific group of IMU errors have been selected that are deemed to be the most important and are
thought to have the largest effect on the integrity of the IMU data. Those are the IMU errors that are
explored in the following subsections.

Section 3.1 – Deterministic Errors

The first category of IMU error is deterministic errors. Deterministic errors can be precisely
determined and represented using either scalar numbers or matrices. Their values either remain constant
over the course of time or vary in a way that can be exactly mathematically modeled. Once a
deterministic error has been accurately modeled, its model will typically remain the same regardless of
how long the IMU is allowed to run or whether power is cycled on the IMU. The following subsections
explain each type of deterministic error in detail and present the mathematical models that are used to
characterize their behavior.

10
Section 3.1.1 – Static Biases

The first type of deterministic error is static biases. A static bias is simply a positive or negative
constant that offsets data by a specific amount. An example of a static bias can be seen in Figure 3.1.
Static biases in IMU data are modeled as a pair of 3x1 vectors as shown below:

= = Eq. 3.1 & 3.2

where represents the static biases for the three-axis accelerometer and represents the static biases
for the three-axis gyroscope. Normally these vectors would be enough to completely model the static
biases. However, due to the temperature sensitive nature of MEMS IMUs, the values of the static biases
change with respect to temperature. The MEMS gyroscopes in particular are more sensitive to
temperature than the MEMS accelerometers, but both sensors must be analyzed for the sake of
completeness. Because of this relationship between the static biases and temperature, the model of the
static biases changes from a single 3x1 vector pair to a series of 3x1 vector pairs arranged in a look-up
table based on temperature. The look-up table takes in a specific input temperature and produces the static
bias vectors that correspond to that temperature.

Figure 3.1: An example of an unbiased sine wave (blue) and a biased sine wave with a positive static bias of 1 (red).

11
Section 3.1.2 – Scale Factors and Misalignment Errors

The second and third types of deterministic errors are scale factors and misalignment errors. Scale
factors will be examined first. Scale factors are similar to static biases in that they offset data by a specific
amount. However, instead of offsetting data by an additive constant, scale factors offset data by a
multiplicative constant. An example of a scale factor can be seen in Figure 3.2. Scale factors in IMU data
are modeled as a pair of 3x3 diagonal matrices as shown below:

0 0 0 0
= 0 0 = 0 0
0 0 0 0
Eq. 3.3 & 3.4

where represents the scale factors for the three-axis accelerometer and represents the scale factors
for the three-axis gyroscope. As was the case with the static biases in Section 3.1.1, scale factors are also
sensitive to temperature. Therefore, instead of being modeled as a single 3x3 matrix pair, scale factors
must be modeled as a series of 3x3 matrix pairs arranged in a look-up table based on temperature. The
look-up table takes in a specific input temperature and produces the scale factor matrices that correspond
to that temperature.
Now misalignment errors will be examined. Misalignment errors occur when the axes of an
IMU’s sensors are not properly aligned with the X, Y and Z directions defined by the IMU package. An

Figure 3.2: An example of a non-scaled sine wave (blue) and a scaled sine wave with a positive scale factor of 0.75 (red).

12
 Figure 3.3: An example of a misalignment between the X direction specified by the IMU package and the X-axis of the IMU
sensors.

example of an axis misalignment can be seen in Figure 3.3.

Since IMUs are typically constructed in square or rectangular formats, the X, Y and Z directions
are usually defined by the unit vectors that are exactly perpendicular to the faces of the IMU package.
Ideally, the axes of the IMU should coincide with the specified X, Y and Z directions exactly. However, it
is very unlikely that this is actually the case. If an axis is not properly aligned with its corresponding
direction, then that axis will measure portions of the accelerations and angular velocities that occur in the
other two directions.
Misalignment errors are particularly noticeable when dealing with acceleration due to gravity.
Suppose that an IMU is sitting on a flat and level lab bench recording data. The axis parallel to the gravity
vector will measure a very large acceleration compared to the other two axes in the plane perpendicular to
the gravity vector (which should each measure close to zero acceleration). If one of those two axes is
misaligned, even by a small amount, it will measure a component of the gravity vector that is proportional
to the amount of misalignment. For example, a misalignment of just 1° with respect to the plane
perpendicular to the gravity vector would cause a measurement of 9.80665 ∗ sin 1° = 0.1711 meters
per second squared in the misaligned axis. This amount of acceleration error may not seem like much of a
problem, but it could be enough to drown out or heavily distort the true acceleration being experienced in
that axis. More importantly, as that acceleration error propagates through the inertial navigation
algorithm, it will cause the estimates of linear velocity, position and orientation to drift significantly
within a matter of minutes or even seconds. In fact, over the course of just five minutes the constant
acceleration error caused by a 1° misalignment would result in a final velocity error of 51.3449 meters per
second and a final position error of 7701.7 meters.
Misalignment errors in IMU data are modeled as a pair of 3x3 matrices as shown below:

13
 = = Eq. 3.5 & 3.6

where represents the axis misalignments in the three-axis accelerometer and represents the axis
misalignments in the three-axis gyroscope. Each of the terms in the and matrices specifies how
accelerations and angular velocities in X, Y and Z directions are measured by the X, Y and Z IMU axes.
For example, the term specifies the component of acceleration in the Y direction that is measured by
the X-axis of the IMU. Ideally, all of the diagonal terms in the and matrices would be equal to one
and all of the other terms would be equal to zero, making both and equal to the identity matrix .
That would mean that the IMU axes are perfectly aligned with the IMU package and each axis is only
ever measuring acceleration and angular velocity along or about its own axis. In reality this is never the
case, but the closer the and matrices are to the identity matrix the smaller the misalignment
errors will be and the better the IMU data will be.
Scale factors and misalignment errors both affect IMU data in very similar ways, and therefore
the mathematical models that define their behavior are also very similar. Because their mathematical
models are so similar, the individual effects of the two errors are difficult to differentiate from one
another. To deal with this issue, the two errors are modeled together as one joint error term. This does not
present a problem in the overall analysis of the IMU data because the effects of the scale factors and
misalignment errors do not need to be known independently. Both of the errors can be compensated for
simultaneously if the combination of their effects is known. The combination of scale factors and
misalignment errors in the IMU data is modeled using a pair of 3x3 matrices as shown below:

= = Eq. 3.7 & 3.8

where represents the combination of the scale factors and misalignment errors in the three-axis
accelerometer and represents the combination of the scale factors and misalignment errors in the three-
axis gyroscope. The temperature sensitive nature of the scale factors described earlier in this section
carries over into the and matrices as well. Therefore, as was the case with the and matrices,
the and matrices must be modeled as a series of 3x3 matrix pairs arranged in a look-up table based
on temperature instead of as a single static 3x3 matrix pair. The look-up table takes in a specific input
temperature and produces the combination of scale factors and misalignment errors that corresponds to
that temperature.

14
Section 3.2 – Stochastic Errors

The second category of IMU error is stochastic errors. Unlike deterministic errors, which are
inherently stable and repeatable, the behavior of stochastic errors is based around random processes. The
random nature of stochastic errors means that their values change constantly over the course of time and
there is no way to predict what the exact value of a specific error term will be at any given point in time.
It also means that if power is cycled on the IMU the values of the stochastic errors will be completely
different from any previous time that the IMU was used. Because of their inherently random nature the
only way to model stochastic errors is by using random variables and probabilistic models. However, it is
assumed that the probability distributions used to model the stochastic errors remain the same over the
course of time. The following subsections explain each type of stochastic error in detail and present the
mathematical models that are used to characterize their behavior.

Section 3.2.1 – Measurement Noise

The first type of stochastic error is measurement noise. Measurement noise is a zero-mean
random process that appears in sensor data and obscures its true nature. An example of measurement
noise can be seen in Figure 3.4. In order to accurately model measurement noise, two properties of the
measurement noise must be understood. The first property is the color of the measurement noise, and the

Figure 3.4: An example of a sine wave without noise (blue) and a sine wave with a white Gaussian noise process (! = 0,
" # = 0.1) added to it (red).

15
second property is the probability density function (PDF) that the measurement noise samples are drawn
from.
The first property used to characterize measurement noise is called the noise color. A noise
process is said to have a certain color based on the relationship between its power spectrum and
frequency. For example, white noise has a power spectrum that is flat or equal valued at all noise
frequencies. Pink noise has a power spectrum that is inversely proportional to frequency (1/%), and
Brownian or random walk noise has a power spectrum that is inversely proportional to the square of the
frequency (1/% #). Some noise colors are relatively easy to model, and some noise colors are not. White
noise and Brownian noise are very simple to model and are also very easy to work with mathematically.
Pink noise, on the other hand, cannot be represented using a finite model and can only ever be
approximated to within a certain degree of accuracy. White noise is one of the more common noise colors
found in both natural and man-made systems. Because of this, noise processes will sometimes be assumed
to be white noise if no other information about the noise process seems to suggest otherwise. However,
the more appropriate method for determining the color of a noise process is to perform a mathematical
analysis such as an Allan Variance or Power Spectral Density (PSD) analysis to get a definitive
characterization of the true color of the noise. These analyses are described in more detail in Sections
4.1.1 and 4.1.2 respectively.
The second property used to characterize measurement noise is called the noise probability
density function (PDF). A PDF is a function that describes how samples from a random noise process are
distributed. As was the case with noise color, there are several different categories of noise PDFs that are
represented by different families of functions. Some examples of PDF families include Gaussian PDFs,
Uniform PDFs, Exponential PDFs, Poisson PDFs, Rayleigh PDFs, and many more. One of the most
commonly used PDF families is the family of Gaussian PDFs (also referred to as Normal PDFs).
Gaussian PDFs are defined by the equation:

1 +, -
% & = *
+
#. -
√2)" #
Eq. 3.9

where ! is the mean of the distribution and " # is the variance of the distribution. The Gaussian PDF has
several properties that make it easier to work with than most PDFs. First, the behavior of many natural
and man-made random processes can be accurately modeled using Gaussian PDFs. Second, a Gaussian
PDF can be completely defined using only two parameters: the mean of the distribution ! and the
variance of the distribution " # . And third, if a linear operation is performed on a Gaussian distributed
random variable, the resulting random variable is also Gaussian distributed. This makes Gaussian

16
distributed random variables very easy to work with in linear systems because any Gaussian distributed
random variable that goes into the system will produce a random variable at the output of the system that
is also Gaussian distributed, but with a potentially different mean and variance.
Because Gaussian distributed noise occurs so frequently and is so desirable mathematically, it is
common practice to initially assume that an unknown noise process is Gaussian distributed if no other
information about the PDF of the noise is available (this assumption can be justified based on the Central
Limit Theorem [9]). However, care must be taken to ensure that this assumption is not made recklessly or
without proper justification. The actual PDF of a noise process should always be verified if possible.
Verifying the PDF of a noise process is described in more detail in Section 4.1.3.

Section 3.2.2 – Turn-On to Turn-On Bias Variation

The second type of stochastic error is turn-on to turn-on bias variation. In Section 3.1.1, static
biases were discussed and classified as a type of deterministic error because their values remain constant
over the course of time. However, that is only true if the static biases are being observed during a single
continuous power cycle. If the static biases are being observed over multiple continuous power cycles,
then there will be slight variations in their values from power cycle to power cycle.
Suppose for example that an IMU is powered on to collect a sample of data, then powered down
after the data collection is complete and then powered on again to perform a second data collection. When
the IMU powers on again to perform the second data collection, the values of the static biases will be
slightly different from the values that were observed during the first data collection. This variation in the
values of the static biases between power cycles is what the turn-on to turn-on bias variation models. The
turn-on to turn-on bias variation for each of the six static biases is modeled as a random process with a
mean and a variance. These variances are used in the initialization of the EKF component of the inertial
navigation algorithm to provide initial covariance values for the bias correction terms in the EKF’s state
vector.

Section 3.2.3 – Drifting Biases

The third type of stochastic error is drifting biases. A drifting bias is a bias that starts off at a
specific value and then randomly drifts away from that value slowly over time (this type of behavior is
commonly referred to as a “random walk” behavior). An example of a drifting bias taken from a sample
of gyroscope data can be seen in Figure 3.5. Unlike static biases, drifting biases contain a random

17
 Figure 3.5: An example of a drifting bias in a sample of IMU gyroscope data.

component that dictates how its value will change over the course of time. This random component must
be taken into account by whatever model is used to approximate the behavior of the drifting bias. The
model chosen to approximate the behavior of the drifting bias is the first-order Markov process model.
The drifting bias model is described by the recursive equation:

/ 0 =1 / 0−1 +4 0 Eq. 3.10

where / 0 is the value of the drifting bias at time 0, 1 is a scale factor, and 4 is a zero-mean white
Gaussian noise (WGN) process with unknown variance "5/
#
. During each iteration of the equation, the
current value of the drifting bias ( / 0 ) is correlated with the previous value ( / 0 − 1 ) through the
scale factor (1), and a sample from the WGN process (4 0 ) is added in to simulate the random
component of the drifting bias. One of the benefits of using the first-order Markov process model to
approximate the behavior of the drifting biases is that because it is a recursive model that is driven by a
WGN process, it is very simple to integrate into the equations of an EKF. In terms of the inertial
navigation algorithm as a whole this means that it is also very simple to compensate for the effects of the
drifting biases, which will improve the overall quality of the linear velocity, position and orientation
estimates that the inertial navigation algorithm produces.

18
Section 4 – Laboratory Experiments

In Section 3, several types of IMU errors were discussed and the mathematical models for
characterizing their behavior were developed. This section describes the laboratory experiments that are
used to determine values for the parameters of the generic IMU error models from Section 3. Each of the
experiments is designed to isolate the effects of one or more of the IMU errors so that their behavior can
be studied independently of the effects of the other IMU errors. Then, by controlling the accelerations and
angular velocities that the IMU experiences during an experiment, the specific error being studied can be
accurately quantified. By the end of this section, each of the IMU errors will have a specific mathematical
model that is tailored to the individual SparkFun and ADI IMUs being used in this thesis.
Each of the following subsections describes a specific experiment in detail, including what the
goal of the experiment is, how the experiment is set up, how the experiment is run, the type of data that is
recorded and how that data is used to calculate the desired IMU error parameters.

Section 4.1 – 12-Hour Static IMU Data Experiment

The goal of the 12-Hour Static IMU Data Experiment is to gather a twelve hour sample of static
IMU data so that the behavior of the measurement noise and drifting biases can be analyzed. Because the
IMU is kept static (meaning that the IMU experiences no accelerations other than those caused by the
Normal force opposing gravity and no angular velocities other than those caused by the rotation of the
Earth) during the duration of the experiment, the only errors that affect the data are measurement noise,
drifting biases, static biases and misalignment errors. In order to isolate the effects of the measurement
noise and drifting biases, it will be assumed that subtracting the average value of the data from the entire
data sample will be enough to sufficiently remove the effects of the static biases and misalignment errors.
The experiment is run for twelve hours to ensure that the drifting biases have enough time to accrue a
significant amount of drift (ten to twelve hours should be enough time for the drifting biases in a typical
IMU to drift off by a noticeable amount). Once the measurement noise and drifting biases have been
isolated, their behaviors can be properly analyzed.
The 12-Hour Static IMU Data Experiment is conducted by following the series of steps in the
bulleted list below:

• The IMU is securely mounted on a test fixture that can be adjusted to ensure that the IMU is
completely level (See Figure 4.1).

19
 Figure 4.1: The test fixture for the 12-Hour Static IMU Data Experiment.

• The test fixture is then enclosed in a hybridization incubator to prevent the internal temperature of
the IMU from changing radically during the course of the test (See Figure 4.2). Because this test
lasts for so long, it is inevitable that the internal temperature of the IMU will change over time as
the temperature of the room changes. Keeping the IMU in the incubator keeps the temperature
fluctuations to a minimum.
• The IMU is then powered on and allowed to come up to temperature (the initial powering on of
the IMU causes its internal temperature to rise approximately 1-2°C).
• Once the IMU has reached a stable temperature, the data logging begins. The data logging lasts
for twelve hours. It is important that the IMU remains completely motionless during this time.
Ideally, the IMU should be left alone in a room by itself with no people around it to ensure that
no unintended vibrations or other disturbances appear in the recorded data.
• Once the data has been recorded, it is filtered to remove outliers and other anomalies. For each
data point (& 6 ) in the sample, the difference between & 6 and the mean value of the sample
(&̅ ) is calculated. If the absolute value of that difference is above a certain threshold (typically 1

& 6 is replaced with the median value of the vector 8& 6 − 1 & 6 + 1 9. If & 6 is the
meter per second squared for the accelerometers and 1 degree per second for the gyroscopes) then
& 6

20
 Figure 4.2: The test fixture for the 12-Hour Static IMU Data Experiment enclosed in a hybridization incubator.

8&̅ & 6 + 1 9 or the median of the vector 8& 6 − 1 &̅ 9 respectively. This
first or the last data point in the sample, then it is replaced with the median value of the vector
& 6 & 6
process removes all of the outliers and other anomalies from the raw inertial data.
• After the data has been filtered, the average value of the recorded data from each individual
sensor axis is subtracted from every point in its respective data sample to remove the effects of
the static biases and misalignment errors. The resulting data samples represent only the effects of
the measurement noise and the drifting biases.

The following subsections describe the various methods that are used to analyze the data gathered
from this experiment.

Section 4.1.1 – Allan Variance Analysis

The first analysis performed on the static IMU data is called an Allan Variance analysis. Allan
Variance was originally developed by Dr. David W. Allan as a method for analyzing the frequency
stability of precision oscillators [10]. However, as mentioned previously in Section 3.2.1, Allan Variance

21
can also be used to analyze a noise process and determine the color or combination of colors that best
models its behavior.

discrete data points & : that is ; elements long and every element is separated by a constant amount of
The method for calculating Allan Variance is as follows. Suppose there exists a time series of

time ∆:. An integer 6 is chosen and & : is divided into segments of 6 data points. If there are any
points left over at the end of & : that don’t fill up an entire segment then they are truncated. A time
= = 6∆: is computed that represents the amount of time spanned by each segment, and the time average
of each of the segments is computed to produce a series of averages denoted by
>? = , ># = … >B = . The averages are then used to compute the Allan Variance " # = for a given
value of = using the equation

B+?
1
" = =
#
C >DE? = − >D = #
2 −1
Eq. 4.1
DF?

This process is then repeated for values of 6 ranging between 1 and ;/10 . The upper limit of ;/10
is imposed because if there are fewer than nine averages used in the calculation of the Allan Variance,
then the results begin to lose their significance [11]. Once the Allan Variance has been calculated over a
sufficient range of 6 values, the square root of the Allan Variance (called the Allan Deviation) is plotted
against the corresponding values of = on a set of log-log axes. This final result is called the Allan
Deviation function.
Once the Allan Deviation function of a noise process has been calculated and plotted, some
insights can be gained about the color of that noise process. Each type of noise color produces a specific
type of Allan Deviation function that is unique to that noise color. For example, the Allan Deviation
function of white noise is a line with a slope of -½ (as seen in Figure 4.3). The Allan Deviation function
of pink noise is a line with a slope of 0, and the Allan Deviation function of Brownian or random walk
noise is a line with a slope of +½. Using this unique relationship between the Allan Deviation function
and noise color, it is possible to identify the color of any noise process by calculating its Allan Deviation
function and comparing it to the Allan Deviation functions of each noise color. Therefore, the color of the
IMU measurement noise can be determined by computing the Allan Deviation function of the static IMU
data (which only contains the effects of the measurement noise and drifting biases) and comparing it to
the Allan Deviation functions of each of the noise colors. The Allan Deviation functions of the static IMU
data can be seen in Figures 4.4 and 4.5.

Initially, when = is small, the Allan Deviation functions are all linearly decreasing. These portions of the
The Allan Deviation functions in Figures 4.4 and 4.5 all have very similar characteristics.

22
 Figure 4.3: The Allan Deviation function of a sample of WGN generated by MATLAB.

Figure 4.4: The Allan Deviation functions of static data from the ADI IMU accelerometers and gyroscopes.

Allan Deviation functions reflect the behavior of the measurement noise. Because of the slow-moving
nature of the drifting biases, when = is small the effects of the measurement noise overpower the effects
of the drifting biases. However, as = gets larger, the Allan Deviation functions all start to flatten out and
curve back up towards a slope value of +½. These portions of the Allan Deviation functions reflect the
behavior of the drifting biases. As = gets larger, the effects of the drifting biases start to become more

23
 Figure 4.5: The Allan Deviation functions of static data from the SparkFun IMU accelerometers and gyroscopes.

Allan Deviation Allan Deviation

ADI IMU SparkFun IMU
Function Slope Function Slope
X-Axis Accelerometer -0.4968 (m/s2)/s X-Axis Accelerometer -0.5023 (m/s2)/s
Y-Axis Accelerometer -0.4941 (m/s2)/s Y-Axis Accelerometer -0.5012 (m/s2)/s
Z-Axis Accelerometer -0.4954 (m/s2)/s Z-Axis Accelerometer -0.4951 (m/s2)/s
X-Axis Gyroscope -0.4980 (°/s)/s X-Axis Gyroscope -0.4977 (°/s)/s
Y-Axis Gyroscope -0.4955 (°/s)/s Y-Axis Gyroscope -0.4832 (°/s)/s
Z-Axis Gyroscope -0.4943 (°/s)/s Z-Axis Gyroscope -0.4978 (°/s)/s

Table 4.1: Measurement noise Allan Deviation function slope values for the ADI and SparkFun IMUs.

prevalent, and eventually they overpower the effects of the measurement noise. Since the ultimate goal of
this analysis is to determine the color of the IMU measurement noise, the later portions of the Allan
Deviation functions will be ignored for now and the rest of the analysis will focus on the early parts of the
Allan Deviation functions where the effects of the measurement noise are the strongest. The later portions
of the Allan Deviation functions will be discussed more in Section 4.1.4.
Upon inspection, the portions of the Allan Deviation functions that reflect the behavior of the
measurement noise are very similar to the Allan Deviation function of white noise. To verify that the
Allan Deviation functions of the measurement noise do truly match up with the Allan Deviation function
of white noise, the value of their slopes must be determined. The calculated slopes of the measurement
noise Allan Deviation functions are listed in Table 4.1.

24
 As Table 4.1 shows, all of the slopes are very close to the value of -½ that is indicative of white
noise. This confirms that the IMU measurement noise can be accurately modeled using the white noise
model.

Section 4.1.2 – Power Spectral Density (PSD) Analysis

The second analysis performed on the static IMU data is called a Power Spectral Density (PSD)
analysis. Like the Allan Variance analysis described earlier, a PSD analysis provides another method for
determining the color of a noise process. PSD analyses are commonly used in the field of signals analysis
to determine the amount of power that a signal has at specific frequencies. Its most basic mathematical
form is defined as the Fourier transform of the autocorrelation function of the signal being analyzed

% = GHI = J Eq. 4.2

where % is the PSD and I = is the autocorrelation function of the signal. However, there are many
variations of PSD and many different ways in which it can be calculated. The method used in this thesis is
called the Welch method. The Welch method was developed by Peter Welch in 1967 as a more
computationally efficient way of estimating the PSD of a signal by calculating the Fourier transforms of
small, overlapping segments within the signal and averaging them together to produce an estimate of the
PSD of the entire signal [12].

Welch’s paper [12]. Suppose there exists a time series of discrete data points K L that is ; elements long
The Welch method is briefly described below, and a more detailed description can be found in

and every element is separated by a constant amount of time ∆:. First, K L is divided up into M fixed-
length, overlapping segments such that the segments cover the entire length of K L as illustrated in
Figure 4.6. Each of the M segments is then filtered by subtracting the average value of the segment from
each of the data points within the segment to remove any biases that may be in the data [13]. Then, each

Figure 4.6: The division of a time series K L into a group of M fixed-length overlapping segments [12].

25
of the M segments is multiplied by a windowing function (any standard windowing function such as
Gaussian, Hamming, Hann, etc. can be used) and its Fourier transform is calculated using the fast Fourier
transform (FFT) algorithm. This results in M Fourier transforms that are then averaged together to
produce an estimate of the PSD of the entire time series K L . The Welch PSD analyses performed in this
thesis were performed using a Gaussian windowing function of size 16384 and an overlap of 50% [12].
As was the case with Allan Variance, there is also a unique relationship between the color of a
noise process and its PSD. These power spectrum to frequency relationships were described earlier in
Section 3.2.1, but they will be briefly restated here for convenience. White noise has a power spectrum that
is flat or equal valued at all noise frequencies (as seen in Figure 4.7), pink noise has a power spectrum that
is inversely proportional to frequency (1/%), and Brownian or random walk noise has a power spectrum
that is inversely proportional to the square of the frequency (1/% #). Using these unique relationships
between PSD and noise color, it is possible to identify the color of any noise process by calculating its PSD
and comparing it to the PSDs of each noise color. Therefore, the color of the IMU measurement noise can
also be determined by taking the PSD of the static IMU data and comparing it to the PSDs of each of the
noise colors. The PSDs of the static IMU data can be seen in Figures 4.8 and 4.9.
As Figures 4.8 and 4.9 show, the PSDs are all essentially flat, which is indicative of white noise.
However, in the range of frequencies less than 1 Hz, the PSDs start to stray away from the flat spectrum
model. This behavior is caused by the drifting biases. As the frequency gets lower the effects of the
drifting biases begin to overpower the effects of the measurement noise and the PSDs start to stray away
from the flat model. But when the frequency gets higher, the effects of the measurement noise begin to

Figure 4.7: The Welch PSD of a sample of WGN generated by MATLAB.

26
 Figure 4.8: The Welch PSDs of static data from the ADI IMU accelerometers and gyroscopes.

Figure 4.9: The Welch PSDs of static data from the SparkFun IMU accelerometers and gyroscopes.

overpower the effects of the drifting biases, and the PSDs fit very closely with the flat model. As was the
case with the Allan Deviation functions, the lower frequency portions of the PSDs will be ignored for
now and the primary focus will be on the portions of the PSDs above the 1Hz frequency.

27
 ADI IMU PSD Slope
X-Axis Accelerometer 8.3417 * 10-6 (m/s2)2/Hz2
Y-Axis Accelerometer 8.5088 * 10-6 (m/s2)2/Hz2
Z-Axis Accelerometer 1.0463 * 10-5 (m/s2)2/Hz2
X-Axis Gyroscope 1.7595 * 10-3 (°/s)2/Hz2
Y-Axis Gyroscope 2.1995 * 10-3 (°/s)2/Hz2
Z-Axis Gyroscope 2.2977 * 10-3 (°/s)2/Hz2
SparkFun IMU PSD Slope
X-Axis Accelerometer 7.4451 * 10-5 (m/s2)2/Hz2
Y-Axis Accelerometer 4.9229 * 10-5 (m/s2)2/Hz2
Z-Axis Accelerometer 1.5559 * 10-4 (m/s2)2/Hz2
X-Axis Gyroscope 1.1222 * 10-4 (°/s)2/Hz2
Y-Axis Gyroscope 1.3046 * 10-4 (°/s)2/Hz2
Z-Axis Gyroscope 1.2000 * 10-4 (°/s)2/Hz2

Table 4.2: Measurement noise PSD slope values for the ADI and SparkFun IMUs.

To verify that the PSDs of the measurement noise do truly match up with the PSD of white noise,
the value of their slopes is calculated. The calculated slopes of the measurement noise PSDs are listed in
Table 4.2.
As Table 4.2 shows, all of the slopes of the measurement noise PSDs are very close to zero,
which indicates that the measurement noise PSDs are all essentially flat. This further confirms that the
IMU measurement noise can be accurately modeled using the white noise model.
Based on the results from both the Allan Variance and PSD analyses, the white noise model is
clearly the model that best fits the IMU measurement noise. Now that the color of the measurement noise
has been determined, all that remains is to determine the PDF of the measurement noise to complete the
measurement noise model.

Section 4.1.3 – Probability Density Function (PDF) Analysis

The third analysis performed on the static IMU data is called a Probability Density Function
(PDF) analysis. The process of determining the PDF of a noise process is relatively simple. The noise
samples are arranged in histograms that display the number of times a specific value appears in the noise
process. Curves can then be fitted to the general shape of these histograms to see which family of PDFs
most accurately models the noise distribution. Figure 4.10 shows the histogram of a sample of WGN with
a Gaussian distribution superimposed on top of it.

28
 Figure 4.10: A histogram of a sample of WGN generated by MATLAB with a Gaussian distribution fitted to it.

Figure 4.11: A histogram of the measurement noise processes from the ADI IMU.

This is the method that will be used to estimate the PDFs of the IMU measurement noise. The
histograms of the IMU measurement noise can be seen in Figures 4.11 and 4.12. The gaps in the
histograms occur because the continuous accelerations and angular velocities being measured by the IMU
are quantized and converted into a digital format by an analog to digital converter (ADC) so that they can

29
 Figure 4.12: A histogram of the measurement noise processes from the SparkFun IMU.

Figure 4.13: The ADI IMU measurement noise histograms with Gaussian distributions fit to them.

be transmitted and processed digitally. The general shape of the histograms seems to suggest that
Gaussian PDFs would accurately model the measurement noise distributions. Because Gaussian
distributions can be completely defined using only a mean and a variance and each of the measurement
noise processes is zero-mean by definition, the sample variance of each measurement noise process can

30
 Figure 4.14: The SparkFun IMU measurement noise histograms with Gaussian distributions fit to them.

ADI IMU Sample Variance SparkFun IMU Sample Variance

X-Axis Accelerometer 1.1267 * 10-4 (m/s2)2 X-Axis Accelerometer 9.3624 * 10-4 (m/s2)2
Y-Axis Accelerometer 1.2580 * 10-4 (m/s2)2 Y-Axis Accelerometer 6.1601 * 10-4 (m/s2)2
Z-Axis Accelerometer 1.4302 * 10-4 (m/s2)2 Z-Axis Accelerometer 2.1651 * 10-3 (m/s2)2
X-Axis Gyroscope 2.2280 * 10-2 (°/s)2 X-Axis Gyroscope 1.4673 * 10-3 (°/s)2
Y-Axis Gyroscope 2.7662 * 10-2 (°/s)2 Y-Axis Gyroscope 5.7326 * 10-3 (°/s)2
Z-Axis Gyroscope 2.9154 * 10-2 (°/s)2 Z-Axis Gyroscope 1.5098 * 10-3 (°/s)2

Table 4.3: The sample variances of the IMU measurement noise processes.

be used to create a family of Gaussian distributions that should accurately model each of the measurement
noises processes.
The sample variances used to compute the Gaussian distributions can be seen in Table 4.3.
Figures 4.13 and 4.14 show the Gaussian distributions superimposed on top of the measurement noise
histograms to provide a visual representation of how accurately they model the measurement noise
distributions. These histograms are only a visual confirmation of the measurement noise PDFs. It could be
proven that the PDFs are Gaussian through hypothesis testing, but visual confirmation is considered to be
enough for this analysis because the Gaussian distributions very closely match the shape of the
measurement noise histograms. Therefore, the PDFs of the IMU measurement noises will be modeled
using Gaussian PDFs parameterized by the sample variances of the measurement noises.

31
 Combining the results obtained from the Allan Variance, PSD and PDF analyses, it is clear that
the IMU measurement noise can be most accurately modeled using the WGN model. This is the model
that will be used to represent the IMU measurement noise in the EKF component of the inertial
navigation algorithm.

Section 4.1.4 – Drifting Bias Analysis

The fourth and final analysis performed on the static IMU data is called the Drifting Bias
analysis. Recall from Section 3.2.2 that the behavior of the drifting bias is modeled using a first-order
Markov process (see Equation 3.10 from Section 3.2.3), which falls under the family of random walk
models. Therefore, the drifting biases should produce Allan Deviation functions with slopes of +½. This
was discussed briefly back in Section 4.1.1, where it was shown that at the tail end of the Allan Deviation

average parameter = gets larger and larger, the effects of the drifting biases start to overpower the effects
functions the slopes started to swing from -½ to +½. This reflects the fact that as the length of the time

of the measurement noise, which is why the effects of the drifting biases show up more predominantly at
the tail ends of the Allan Deviation functions.
The process of estimating the parameters for the first-order Markov processes that characterize
the drifting biases begins by first isolating the drifting biases from a sample of static IMU data. As
mentioned previously, the sample of static IMU data should be at least twelve hours in length to ensure
that the drifting biases have an adequate amount of time to diverge from their initial values. The drifting
biases can be isolated from a sample of static IMU data by first subtracting the sample means of the data
samples from each individual sensor axis to remove the effects of the static biases and then performing a
moving average on the data to remove the effects of the measurement noise. The remaining portions of
the IMU data represent the drifting biases.
The parameters of the first-order Markov process can be estimated by exploiting a couple of
mathematical properties of the first-order Markov process model. There are two parameters that must be
estimated for each drifting bias: a scale factor 1 and a noise variance "5/
#
that is the variance of the WGN
process that drives the first-order Markov process. The parameter 1 can be found by analyzing the
autocorrelation function (ACF) of the drifting bias. The value of the ACF of a first-order Markov process
at a lag value (=) equal to one represents the amount of correlation that exists between any two data points
in the first-order Markov process that are separated by one timestep. Referring back to the discussion of
the drifting biases in Section 3.2.3, the parameter 1 was also defined as the amount of correlation that
exists between any two data points in a drifting bias that are separated by one timestep. Therefore, since

32
 ADI IMU P QTRS
X-Axis Accelerometer 0.9887 1.9239 * 10-7 (m/s2)2
Y-Axis Accelerometer 0.9912 3.5379 * 10-7 (m/s2)2
Z-Axis Accelerometer 0.9907 2.4626 * 10-7 (m/s2)2
X-Axis Gyroscope 0.9923 4.3040 * 10-6 (°/s)2
Y-Axis Gyroscope 0.9789 8.2511 * 10-6 (°/s)2
8.1647 * 10-6 (°/s)2
P QTRS
Z-Axis Gyroscope 0.9908
SparkFun IMU
X-Axis Accelerometer 0.9907 1.0827 * 10-7 (m/s2)2
Y-Axis Accelerometer 0.9905 3.8738 * 10-9 (m/s2)2
Z-Axis Accelerometer 0.9907 3.9915 * 10-6 (m/s2)2
X-Axis Gyroscope 0.9907 1.1606 * 10-6 (°/s)2
Y-Axis Gyroscope 0.9809 1.5477 * 10-4 (°/s)2
Z-Axis Gyroscope 0.9906 2.1167 * 10-7 (°/s)2

Table 4.4: The estimated values of the drifting bias first-order Markov process parameters.

these two values represent the same statistical quantity, the value of 1 for each drifting bias can be
estimated by simply calculating the ACF of the drifting bias data and using the value of the ACF at a lag
value of one [14].
Once the 1 parameter has been estimated, there is another mathematical property of the first-
order Markov process that can be exploited to produce an estimate of the "5/
#
parameter. The variance of
a first-order Markov process is defined as

"5/
#
N>O =
/
1 − 1#
Eq. 4.3

where N>O / is the variance of the first-order Markov process (in this case the drifting bias). Therefore,
an estimate of the noise variance "5/
#
can be produced by taking the variance of the drifting bias data and
multiplying it by the quantity 1 − 1 # . Now that both the 1 and "5/
#
parameters have been estimated, the
first-order Markov process that represents the behavior of the drifting bias is complete. Table 4.4 shows
the values that were obtained for all of the drifting bias first-order Markov process parameters [14].
The results displayed in Table 4.4 show that on average the drifting biases in the gyroscopes tend
to drift more noticeably than drifting biases in the accelerometers do. The results also show that the values
of 1 for both IMUs are very close to one and the noise variances are all very small. These values suggest
that the amount of bias drift is fairly minimal, and it shouldn’t have a large impact on the overall integrity
of the raw inertial data.

33
Section 4.2 – Power Cycle Experiment

The goal of the Power Cycle Experiment is to determine the amount of variance in the values of
the accelerometer and gyroscope static biases. This experiment is similar to the 12-Hour Static IMU Data
Experiment, except the length of the data sample will be much shorter (approximately five minutes) and a
group of static data samples will be taken in between power cycles of the IMU in order to determine the
turn-on to turn-on bias variance. A power cycle is defined as the act of turning the IMU off and then back
on again at a later time. As was the case with the 12-Hour Static IMU Data Experiment, the only errors
that affect the data during data collection are measurement noise, drifting biases, static biases and
misalignment errors. In order to isolate the effects of the static biases, each data sample is limited to a
length of five minutes, over which time the effects of the drifting biases are considered to be negligible.
Each data sample is then averaged together to produce a single set of average values for the data sample.
The measurement noise in the data cannot be perfectly removed, but its effects are dramatically reduced
to the point of being negligible (i.e. variances on the order of 10-7 to 10-8 (m/s2)2 for the accelerometers
and 10-6 to 10-7 (°/s)2 for the gyroscopes) once the data sample is averaged together because the
measurement noise is zero-mean. The effects of the misalignment errors are also considered to be
negligible for the purposes of this experiment. Once the static biases have been isolated, their behavior in

Figure 4.15: The test fixture for the Power Cycle Experiment.

34
 Figure 4.16: The test fixture for the Power Cycle Experiment enclosed in a hybridization incubator.

between power cycles of the IMU can be properly analyzed.

The Power Cycle Experiment is conducted by following the series of steps in the bulleted list
below:

• The IMU is securely mounted on a test fixture that can be adjusted to ensure that the IMU is
completely level (See Figure 4.15). The test fixture is then enclosed in a hybridization incubator
to minimize fluctuations in the internal temperature of the IMU during the course of the test (See
Figure 4.16).
• The IMU is then powered on and allowed to come up to temperature (the initial powering on of
the IMU causes its internal temperature to rise approximately 1-2°C).
• Once the IMU has reached a stable temperature, a five minute data sample is taken. It is
important that the IMU remains completely motionless during this time.
• After the data sample is taken, the IMU is unplugged and powered off for five minutes. After five
minutes, the IMU is plugged back in and powered on. The IMU is allowed to come up to
temperature again and once the internal temperature of the IMU has stabilized, another five
minute data sample is taken. Then the IMU is unplugged and powered off for another five

35
 ADI IMU Turn-On to Turn-On Bias Variation
X-Axis Accelerometer 6.7203 * 10-6 (m/s2)2
Y-Axis Accelerometer 8.7258 * 10-6 (m/s2)2
Z-Axis Accelerometer 4.2737 * 10-6 (m/s2)2
X-Axis Gyroscope 7.2805 * 10-5 (°/s)2
Y-Axis Gyroscope 1.9517 * 10-4 (°/s)2
Z-Axis Gyroscope 4.4230 * 10-4 (°/s)2
SparkFun IMU Turn-On to Turn-On Bias Variation
X-Axis Accelerometer 1.1761 * 10-6 (m/s2)2
Y-Axis Accelerometer 1.5348 * 10-6 (m/s2)2
Z-Axis Accelerometer 4.3348 * 10-5 (m/s2)2
X-Axis Gyroscope 2.1586 * 10-4 (°/s)2
Y-Axis Gyroscope 2.4880 * 10-3 (°/s)2
Z-Axis Gyroscope 5.1009 * 10-5 (°/s)2

Table 4.5: The calculated values of the turn-on to turn-on bias variation for the ADI and SparkFun IMUs.

minutes before the next data sample is taken. This process repeats until twenty distinct data
samples have been collected.
• Once the data has been recorded, it is filtered to remove outliers and other anomalies (see Section
4.1 for a more detailed description of the filtering process). Then the inertial data in each
individual data sample is averaged together to dramatically reduce the effects of the measurement
noise. The resulting averaged data sets represent only the static biases of each of the individual
sensor axes. The twenty data points for each individual sensor axis are then grouped together and
the sample variances of those twenty data points are calculated to determine the turn-on to turn-
on bias variation for each of the individual sensor axes.
After performing the Power Cycle Experiment in the manner described above, specific values for
the turn-on to turn-on bias variations for both the ADI and SparkFun IMUs were calculated. Table 4.5
shows the results.

Section 4.3 – Multi-Angle Experiment

The goal of the Multi-Angle Experiment is to determine the values of the static biases, scale
factors and misalignment errors for an IMU’s accelerometer as well as the extent to which those values
are affected by temperature. This experiment is based on the fact that gravity is a known source of
constant acceleration. By fixing the IMU in a known orientation it is possible to calculate the amount of
acceleration due to gravity that each IMU axis should experience. These ideal accelerations can then be

36
 Figure 4.17: The Multi-Angle Experiment test fixture with the IMU mounted to it in both configurations.

Figure 4.18: The Multi-Angle Experiment test fixture enclosed in the hybridization incubator. The PVC pipes are used to
ensure that the test fixture remains stationary and level during the course of the experiment.

compared to the measured accelerations produced by the IMU and based on that comparison the values of
the static biases, scale factors and misalignment errors for the accelerometer can be estimated.
Due to the temperature sensitive nature of the static biases and scale factors, this experiment must
be conducted over a range of different temperatures. A look-up table can then be created using the results
from each temperature. Since the ADI IMU is already temperature calibrated by the manufacturer, only
the SparkFun IMU is tested over multiple temperatures. The static biases, scale factors and misalignment
errors for the ADI IMU accelerometer will be represented by a single set of parameters.
The Multi-Angle Experiment is conducted by following the series of steps in the bulleted list
below:

37
• Before conducting the experiment, an appropriate test fixture must first be created. An image of
the test fixture can be seen in Figure 4.17. The test fixture consists of a fixed wooden frame with
a rotating arm mounted to it such that the axis of rotation is perpendicular to the gravity vector.
The IMU is rigidly mounted to the rotating arm such that one of its axes is aligned with the test
fixture’s axis of rotation and the other two axes are exposed to gravity in varying degrees based
on the rotating arm’s orientation. The test fixture has several pre-drilled holes in it that mark
increments of 45° all the way around the test fixture. There is a screw embedded in the rotating
arm that can be screwed into these holes, which allows the rotating arm to be fixed at increments
of 45° all the way around the test fixture.
• In order to determine how sensitive the static biases and scale factors are to temperature, data will
be collected from an array of different temperatures (Room Temperature, 35°C, 40°C, 45°C,
50°C, 55°C and 60°C). To maintain these target temperatures during the course of the
experiment, the test fixture is enclosed in a hybridization incubator (as shown in Figure 4.18). A
digital level is used to make sure that the fixture is level inside the incubator and that its axis of
rotation is as perpendicular to gravity as possible.
• Once the test fixture has been placed inside the incubator, the orientation of the IMU with respect
to gravity is measured using the digital level and recorded.
• The incubator is then allowed to saturate at the first target temperature. Once the incubator has
saturated, the IMU is powered on and its internal temperature is allowed to saturate at the target
temperature.
• After the IMU has come up to temperature, a five minute static data sample is taken. After the
data sample has been collected, the IMU is unplugged and the incubator is set for the next target
temperature. Once the incubator saturates at the new target temperature, the IMU is powered on
and allowed to saturate at the new target temperature. Once the IMU has come up to temperature,
another five minute static data sample is taken. Then the IMU is unplugged, the incubator is set
for the next target temperature and the process repeats until all of the target temperatures have
been covered.
• After covering all of the target temperatures, the test fixture is rotated to the next 45° increment.
The orientation of the IMU is measured again using the digital level and recorded. The data
logging procedure is then repeated at each of the target temperatures with the IMU in its new
orientation.
• This process is repeated until all eight orientations on the test fixture have been tested. Then the
IMU is removed from the test fixture and remounted such that the previously untested IMU axis
(the axis in-line with the rotation of the test fixture) will be exposed to gravity. The eight

38
 orientations of the test fixture are tested again in the same manner as before, resulting in a total of
sixteen orientations.
• After the experiment is over, each of the five minute data samples is filtered to remove outliers
and other anomalies (see Section 4.1 for a more detailed description of the filtering process).
Then the average of each data sample is calculated to produce the average X, Y and Z
acceleration for each data sample. Because the data samples are only five minutes long, the
effects of the drifting biases are considered to be negligible. However, since the IMU is power
cycled in between data samples, turn-on to turn-on bias variation must be accounted for in the
data analysis. The measurement noise in the data cannot be perfectly removed, but its effects are
dramatically reduced to the point of being negligible (i.e. variances on the order of 10-7 to 10-8
(m/s2)2 for the accelerometers and 10-6 to 10-7 (°/s)2 for the gyroscopes) once the data sample is
averaged together because the measurement noise is zero-mean.
• The final result is that for each target temperature there should be an average X, Y and Z
acceleration for each of the sixteen distinct IMU orientations. This data can then be processed
using a least squares estimation algorithm to determine the static biases, scale factors and
misalignment errors for the accelerometer at each of the target temperatures. A lookup table can
then be created by interpolating between the target temperatures to determine the static biases,
scale factors and misalignment errors for any given input temperature.

Section 4.3.1 – Linear Least Squares Estimation Analysis

For each target temperature, the Multi-Angle Experiment produces an average measured X, Y
and Z acceleration for each of the sixteen IMU orientations. Also, because the actual orientation of the
IMU at each orientation set point was recorded using a digital level, the actual accelerations that were
experienced by the IMU at each of the sixteen orientations can be calculated as well. These values
represent the true accelerations that would have ideally been measured by the accelerometer if it could
produce perfect measurements. To estimate the values of the static biases, scale factors and misalignment
errors for the accelerometer, the measured accelerations will be compared to the true accelerations
through a process called linear least squares estimation (LLSE). LLSE is an estimation technique that
applies to linear systems with Gaussian distributed noise in the form of

U = V& + N Eq. 4.4

39
where & is some unknown state vector to be estimated, V is a known observation matrix that performs an
operation on the & parameters, U is a vector of measurements made of the & parameters after being
operated on by V, and N is a zero-mean WGN process with a covariance matrix I that obscures the
measurement data U. LLSE provides a method for optimally estimating (in terms of minimum error
variance) the value of & based on the imperfect measurements U. This optimal estimate can be produced
by using the equation

&W = V X I +? V +?
VX I+? U Eq. 4.5

where &W is the optimal estimate of the state vector & [15].
These equations can be interpreted to fit the data from the Multi-Angle Experiment and produce
an estimate of the static biases, scale factors and misalignment errors for the accelerometer. The
relationship between the measured accelerations and true accelerations, incorporating the effects of the
static biases, scale factors and misalignment errors, is

>B = >Y + +Z Eq. 4.6

where >B is the measured acceleration, >Y is the true acceleration, and Z is the IMU measurement noise.
This equation can be rearranged to fit the form of Equation 4.4. The averaged X, Y and Z acceleration
measurements are used to populate the U vector, the true X, Y and Z acceleration values are used to
populate the V matrix, the static bias, scale factor and misalignment error parameters to be estimated are
used to populate the & vector and the IMU measurement noise is used to populate the N vector. The new
linear vector equation is

_ c
^ b
^ b
^` b
^` b
>B 1 0 0 >Y >Y >Y 0 0 0 0 0 0 ^ b Z
`
[>B \ = 0 1 0 0 0 0 >Y >Y >Y 0 0 0 ^ b+ Z
`
>B 0 0 1 0 0 0 0 0 0 >Y >Y >Y ^ b Z
Eq. 4.7

^` b
^` b
^ b
^` b
^` b
]` a

40
 U? = V? & + N? Eq. 4.8

This equation is mathematically the same as Equation 4.6, but it is in a form that is more conducive to
LLSE. Equation 4.8, however, is only enough to represent data from a single orientation. In order to use
the data from all sixteen orientations in the LLSE algorithm, sixteen of these matrix equations are created,
each one using the data from one of the individual IMU orientations. The individual equations are then
stacked together to produce one combined matrix equation as shown in Equation 4.9.

U? V? N?
U# V# N
⋮ = ⋮ & + ⋮ → U = V& + N
#

U?e V?e N?e

Eq. 4.9

The last matrix that needs to be populated is the covariance matrix I. The I matrix is populated

Rg hg
−0.0259 1.0002 −0.0085 −0.0008
ADI IMU

[−0.0763\ (m/s2) [ 0.0100 1.0003 −0.0048\

0.0148 −0.0005 −0.0005 1.0001
Room Temperature

Rg hg
0.2663 1.0529 −0.0160 0.0040
SparkFun IMU

[−0.1030\ (m/s2) [0.0124 1.0630 0.0093\

−0.6025 0.0039 −0.0051 1.0167
Room Temperature

0.2211 1.0521 −0.0175 0.0030

[−0.1137\ (m/s2) [0.0139 1.0665 0.0100\
−0.5437 0.0047 −0.0054 1.0186
35°C

0.2213 1.0528 −0.0176 0.0031

[−0.1248\ (m/s2) [0.0141 1.0673 0.0102\
−0.5848 0.0048 −0.0062 1.0174
40°C

0.2258 1.0539 −0.0176 0.0034

[−0.1419\ (m/s2) [0.0139 1.0684 0.0109\
0.6261 0.0034 −0.0065 1.0172
45°C

0.2316 1.0546 −0.0177 0.0034

[−0.1561\ (m/s2) [0.0140 1.0691 0.0112\
−0.6583 0.0038 −0.0076 1.0154
50°C

0.2383 1.0556 −0.0178 0.0036

[−0.1723\ (m/s2) [0.0137 1.0700 0.0121\
−0.6905 0.0027 −0.0077 1.0154
55°C

0.2427 1.0564 −0.0178 0.0041

[−0.1887\ (m/s2) [0.0138 1.0706 0.0123\
−0.7264 0.0027 −0.0084 1.0135
60°C

Table 4.6: The estimated values of the static biases, scale factors and misalignment errors for the ADI and SparkFun IMU
accelerometers.

41
by repeating the three accelerometer turn-on to turn-on bias variances (discussed in Section 4.2) along the
diagonal of the matrix as shown in Equations 4.10 and 4.11. The I matrix is populated with the turn-on to
turn-on bias variances because the IMU is powered cycled in between individual data samples. The turn-
on to turn-on bias variances account for the slight variations in the accelerometer static biases that occur
over multiple independent data samples as a result of this power cycling.

"# 0 0 I′ ⋯ 0
I′ = 0 "# 0 I=[⋮ ⋱ ⋮\
0 0 "# 0 ⋯ I′ no
Eq. 4.10 & 4.11
no

Now that all of the required matrices are known, they can be plugged into Equation 4.5 to
produce estimates of the static biases, scale factors and misalignment errors for the accelerometer.
After performing the Multi-Angle Experiment in the manner described above, specific values for
the static biases, scale factors and misalignment errors for both the ADI and SparkFun accelerometers
were estimated for all of the target temperatures. Table 4.6 shows the results.

Section 4.4 – Rate Table Experiment

The goal of the Rate Table Experiment is to determine the values of the static biases, scale factors
and misalignment errors for an IMU’s gyroscope as well as the extent to which those values are affected
by temperature. This experiment is very similar in design to the Multi-Angle Experiment described in
Section 4.3. In the Multi-Angle Experiment, gravity was used as a known source of constant acceleration
to determine the static biases, scale factors and misalignment errors for an accelerometer. Unfortunately,
there is no convenient naturally occurring source of constant angular velocity that can be exploited in this
experiment to determine the static biases, scale factors and misalignment errors for a gyroscope.
Therefore, a stepper motor is used to create a known and controllable angular velocity for the gyroscope
to measure. By mounting the IMU to the stepper motor and rotating it at a known and constant angular
velocity, it is possible to calculate the amount of angular velocity that each IMU axis should experience.
These ideal angular velocities can then be compared to the measured angular velocities produced by the
IMU and based on that comparison the values of the static biases, scale factors and misalignment errors
for the gyroscope can be estimated.
As was the case with the Multi-Angle Experiment, the temperature sensitive nature of the static
biases and scale factors means that the Rate Table Experiment must also be conducted over a range of
different temperatures. A look-up table can then be created using the results from each temperature.

42
 Figure 4.19: The Rate Table Experiment test fixture with the IMU mounted to it in all three configurations.

Again, since the ADI IMU is already temperature calibrated, only the SparkFun IMU is tested over
multiple temperatures. The static biases, scale factors and misalignment errors for the ADI IMU
gyroscope will be represented by a single set of parameters.
The Rate Table Experiment is conducted by following the series of steps in the bulleted list
below:

• Before conducting the experiment, an appropriate test fixture must first be created. An image of
the test fixture can be seen in Figure 4.19. The test fixture consists of a stepper motor with a
rotating platform attached to the shaft via a shaft collar. The IMU is mounted to the rotating
platform such that one of its axes is in line with the axis of rotation of the stepper motor. The
stepper motor is controlled using a microcontroller, which sends digital signals to a stepper motor
controller at a constant frequency. These signals determine when the motor should step and in
what direction. The frequency of the signals coming from the microcontroller determines the
angular velocity of the stepper motor. The stepper motor is mounted to a wooden base to ensure
that it doesn’t move during the duration of the testing. The wooden base has feet that can be
adjusted to ensure that the stepper motor platform is level and the axis of rotation of the stepper
motor is parallel to gravity. The test fixture is powered using an external 12V power source to
ensure that the stepper motor has enough power to rotate at a constant velocity.
• Unlike many of the other experiments performed in this thesis where the IMU remains stationary
for the duration of the experiment, in this experiment the IMU is constantly rotating. This means
that data from the IMU cannot be logged onto a computer using a direct USB connection because
the USB cable would very quickly become tangled around the test fixture and could potentially
damage the test fixture, the IMU or itself. To overcome this problem the IMU was outfitted with
a 3.3V external lithium polymer battery to supply it with power and a data logger that reads the
inertial data coming out of the IMU over its serial port and records the data directly into a text file

43
 that is stored on an on-board 2GB microSD card (see Figure 4.20). The microSD card can then be
removed after the experiment has concluded and the data from the experiment can be transferred
from the microSD card to a computer.
• In order to determine how sensitive the static biases and scale factors are to temperature, data will
be collected from an array of temperatures (Room Temperature, 35°C, 40°C, 45°C and 50°C
because the lithium polymer battery used in this experiment is only rated for temperatures up to
50°C). To maintain these target temperatures during the course of the experiment, the test fixture
is enclosed in a hybridization incubator (as shown in Figure 4.21). A digital level is used to make
sure that the test fixture is level inside the incubator and that the axis of rotation of the stepper
motor is as parallel to gravity as possible. Since there is no direct USB connection between the
IMU and a computer during this experiment, monitoring the internal temperature of the IMU is
more difficult. To overcome this problem, a second identical IMU is placed inside the incubator
to act as a stationary temperature reference. The stationary IMU is connected to a computer with
a direct USB connection and its internal temperature is observed in real time during the course of
the experiment. Because the two IMUs are identical, the unknown internal temperature of the
rotating IMU can be inferred from the known internal temperature of the stationary IMU.
• The incubator is then allowed to saturate at the first target temperature. Once the incubator has

Figure 4.20: The IMU data logging peripherals.

44
 Figure 4.21: The Rate Table Experiment test fixture enclosed in the hybridization incubator.

saturated, the IMU is powered on and its internal temperature is allowed to saturate at the target
temperature.
• After the IMU has come up to temperature, the stepper motor circuitry is powered on and the
IMU begins to rotate at a specific angular velocity. A five minute data sample is taken while the
IMU is rotating. After the data sample has been collected the stepper motor circuitry is powered
off, the IMU is unplugged and the incubator is set for the next target temperature. Once the
incubator saturates at the new target temperature, the IMU is powered on and allowed to saturate
at the new target temperature. Once the IMU has come up to temperature, the stepper motor
circuitry is powered on again and another five minute data sample is taken while the IMU is
rotating. Then the stepper motor circuitry is powered off, the IMU is unplugged, the incubator is
set for the next target temperature and the process repeats until all of the target temperatures have
been covered.
• After covering all of the target temperatures, the data logging procedure is repeated at each of the
target temperatures again using a different angular velocity for the stepper motor. This process is
repeated for five different angular velocities (0 degrees per second, ±90 degrees per second and
±180 degrees per second). Then the IMU is removed from the test fixture and remounted such

45
 that one of the previously untested IMU axes (one of the axes perpendicular to the stepper
motor’s axis of rotation) will be exposed to the rotation of the stepper motor. The five angular
velocities are tested again in the same manner as before for the two remaining IMU axes,
resulting in a total of fifteen angular velocities.
• After the experiment is over, each of the five minute data samples is filtered to remove outliers
and other anomalies (see Section 4.1 for a more detailed description of the filtering process).
Then the average of each data sample is calculated to produce the average X, Y and Z angular
velocity for each data sample. Because the data samples are only five minutes long, the effects of
the drifting biases are considered to be negligible. However, since the IMU is power cycled in
between data samples, turn-on to turn-on bias variation must be accounted for in the data
analysis. The measurement noise in the data cannot be perfectly removed, but its effects are
dramatically reduced to the point of being negligible (i.e. variances on the order of 10-7 to 10-8
(m/s2)2 for the accelerometers and 10-6 to 10-7 (°/s)2 for the gyroscopes) once the data sample is
averaged together because the measurement noise is zero-mean.
• The final result is that for each target temperature there should be an average X, Y and Z angular
velocity for each of the fifteen distinct IMU configurations with the stepper motor. This data can
be processed using the same least squares estimation algorithm that was used in the Multi-Angle
Experiment to determine the static biases, scale factors and misalignment errors for the gyroscope
at each of the target temperatures. A lookup table can then be created by interpolating between
the target temperatures to determine the static biases, scale factors and misalignment errors for
any given input temperature.

Section 4.4.1 – Linear Least Squares Estimation Analysis

For each target temperature, the Rate Table Experiment produces an average measured X, Y and
Z angular velocity for each of the fifteen IMU configurations with the stepper motor. Also, the actual
angular velocities that were experienced by the IMU in each of the fifteen configurations are known.
These values represent the true angular velocities that would have ideally been measured by the
gyroscope if it could produce perfect measurements. To estimate the values of the static biases, scale
factors and misalignment errors for the gyroscope, the measured angular velocities will be compared to
the true angular velocities using the same LLSE algorithm that was used to analyze the data from the
Multi-Angle Experiment (see Section 4.2.1 for more information on how the LLSE algorithm works and
how to convert the data into the proper matrices that are required by the algorithm). The parameters for
Equations 4.6 and 4.7 remain the same with the exception that angular velocities are used in place of

46
 Rp hp
−0.0746 1.0010 −0.0007 −0.0005
ADI IMU

[ 0.0325 \ (°/s) [ 0.0017 1.0006 −0.0117\

−0.2106 −0.0057 −0.0218 1.0039
Room Temperature

Rp hp
−0.9049 0.9849 −0.0059 −0.0159
SparkFun IMU

[ 0.0824 \ (°/s) [0.0057 0.9897 0.0043 \

1.3259 0.0047 0.0152 1.0008
Room Temperature

−1.0624 0.9908 0.0130 −0.0033

[ 1.4117 \ (°/s) [0.0051 0.9925 0.0014 \
1.3279 0.0115 0.0102 0.9979
35°C

−1.1308 0.9907 0.0129 −0.0032

[ 2.2828 \ (°/s) [0.0051 0.9926 0.0012 \
1.3470 0.0112 0.0102 1.0019
40°C

−1.2068 0.9902 0.0125 −0.0031

[ 3.3385 \ (°/s) [0.0045 0.9894 0.0010 \
1.3662 0.0106 0.0104 1.0018
45°C

−1.2657 0.9902 0.0125 −0.0031

[ 4.1985 \ (°/s) [0.0046 0.9887 0.0009 \
1.3921 0.0107 0.0103 0.9995
50°C

−1.3078
[ 5.0648 \ (°/s)
1.4165
55°C *N/A

−1.3621
[ 6.0811 \ (°/s)
1.4481
60°C *N/A

*NOTE: Gyroscope scale factor matrices ( ) could not be calculated for 55°C and 60°C because the lithium polymer
battery used for the Rate Table Experiment was only rated for temperatures up to 50°C. The gyroscope bias matrices ( ) for
the 55°C and 60°C temperatures were calculated from the corresponding Multi-Angle Experiment data sets.

Table 4.7: The estimated values of the static biases, scale factors and misalignment errors for the ADI and SparkFun IMU
gyroscopes.

accelerations. Once all of the appropriate matrices have been calculated and organized into the proper
format (as dictated by Equations 4.8 and 4.9), they can be plugged into Equation 4.5 to produce an
estimate of the static biases, scale factors and misalignment errors for the gyroscope.
After performing the Rate Table Experiment in the manner described above, specific values for
the static biases, scale factors and misalignment errors for both the ADI and SparkFun gyroscopes were
estimated for all of the target temperatures. Table 4.7 shows the results.

47
Section 5 – The Inertial Navigation Algorithm

Now that all of the IMU error sources have been properly analyzed and quantified, the next step
is to develop the inertial navigation algorithm that will produce estimates of the linear velocity, position
and orientation of a vehicle based on the raw IMU data and the secondary navigation information. The
following subsections provide a basic overview of the entire inertial navigation algorithm as a whole, and
more detailed descriptions of the individual components of the inertial navigation algorithm as well.

Section 5.1 – Coordinate Frames

The first step in understanding the inertial navigation algorithm is to understand the various
coordinate frames that are used to describe how specific pieces of information are interpreted. The inertial
navigation algorithm presented in this thesis uses four coordinate frames: the Body frame, the Navigation
frame, the Earth frame and the Inertial frame. Figure 5.1 shows the four coordinate frames and how they
are related to each other.
Each of the four coordinate frames used by the inertial navigation algorithm will now be
explained in more detail. The first coordinate frame in the chain is the Body frame. The origin of the
Body frame is fixed to the center of the vehicle (or another more applicable reference point) and its axes
are oriented such that the X-axis points out through the front of the vehicle, the Z-axis points down
through the bottom of the vehicle and the Y-axis points out through the right side of the vehicle. The
Body frame translates and rotates with the vehicle as it moves through three-dimensional space, and the
origin of the Body frame is used as the reference point for the (x, y, z) position of the vehicle. The Body

Figure 5.1: The four coordinate frames used by the inertial navigation algorithm.

48
frame is also used to define the positive acceleration and angular velocity conventions for the system.
Positive accelerations are defined as accelerations that occur along the positive axes of the Body frame,
and positive angular velocities are defined as shown in Figure 5.2.
The next coordinate frame up the chain from the Body frame is the Navigation frame. The origin
of the Navigation frame is initially set at the starting location of the vehicle and its axes are oriented such
that the X-axis points due North, the Z-axis points straight down towards the center of the Earth and the
Y-axis points due East (this is also commonly referred to as the North, East, Down or NED convention).
Unlike the Body frame, the Navigation frame does not rotate with the vehicle. Its orientation is
permanently fixed according to the NED convention. The Navigation frame may or may not translate with
the vehicle depending on how far the vehicle travels. In order to fully understand how the Navigation
frame is defined, the next coordinate frame up the chain must first be introduced.
The next coordinate frame up the chain from the Navigation frame is the Earth frame. The origin
of the Earth frame is fixed at the center of the Earth and its axes are oriented such that the X-axis points
out of the Earth along the prime meridian, the Z-axis points out of the Earth through the North Pole and
the Y-axis is aligned to complete the right-handed orthogonal coordinate system. The Earth frame rotates
about its Z-axis along with the Earth (at a rate of approximately fifteen degrees per hour) so that the X-
axis of the Earth frame is always aligned with the Prime Meridian.
In navigation applications where the vehicle is navigating over a small area, the origin of the

Figure 5.2: The positive angular velocity conventions defined by the Body frame.

49
Navigation frame remains permanently fixed at the same location on the Earth’s surface (the starting
location of the vehicle) and does not translate with the vehicle. If the vehicle is navigating over a small
enough area, then the curvature of the Earth’s surface is so slight that it can be considered negligible.
Because the curvature of the Earth’s surface is negligible in this scenario, the navigation of the vehicle
can be constrained to a single planar surface instead of a curved surface. This surface is called the Local
Tangent Plane (LTP), and it is defined as the plane created by the X- and Y-axes of the Navigation frame.
The LTP is a perfectly flat surface that is tangent to the surface of the Earth at a single point (the starting
location of the vehicle). Using the LTP greatly simplifies many aspects of the inertial navigation
algorithm, but the downside is that it is only accurate over small areas. Using the LTP to define the
position of the vehicle becomes more and more inaccurate as the vehicle travels further and further away
from its starting location.
In navigation applications where the vehicle is navigating over a large area (i.e. navigating on a
more global scale like an airplane would), the origin of the Navigation frame is moved to coincide with
the origin of the Body frame and it translates with the vehicle as it moves across the Earth’s surface. In
these global-scale navigation scenarios, the curvature of the Earth becomes non-negligible and therefore
the assumptions made by the LTP become unrealistic. The vehicle must be modeled as navigating on the
curved surface of the Earth instead of the planar surface of the LTP. This means that the position of the
vehicle must be represented in terms of spherical coordinates (latitude, longitude, altitude) instead of
planar coordinates (NED). Although, because the Earth is not perfectly spherical in shape, an oblate
spheroid model of the Earth is used to more accurately model the true shape of the Earth. The Global
Positioning System (GPS) uses the WGS84 model of the Earth to produce more accurate measurements
of the latitude, longitude and altitude of a GPS receiver on the surface of the Earth.
The last coordinate frame in the chain is the Inertial frame. The Inertial frame is a coordinate
frame whose origin is fixed at some point in outer space far away from the surface of the Earth. For the
purposes of this thesis, the inertial frame is simply treated as a reference frame that can be used to observe
the rotation of the Earth.

Section 5.2 – Assumptions

This subsection lists the assumptions that are made by the inertial navigation algorithm presented
in this thesis. Each of the assumptions is described in the bulleted list below.

• It is assumed that the vehicle is always navigating over a small enough area that the LTP
assumptions are always valid (i.e. the origin of the Navigation frame remains fixed at a single

50
 point on the Earth’s surface (the starting location of the vehicle) and doesn’t translate with the
vehicle). Therefore, the position of the vehicle is referenced to the Navigation frame as opposed
to the Earth frame and is denoted using NED planar position coordinates.
• It is assumed that the inertial data returned by the IMU conforms to the conventions of positive
acceleration and positive angular velocity defined by the Body frame (seen Section 5.1 for
details). If this is not the case, then the IMU data must be appropriately transformed to match the
positive acceleration and positive angular velocity conventions defined by the Body frame. It is
also assumed that any external misalignments between the IMU package and the vehicle are
small enough that their effects can be considered negligible. Therefore, the orientation of the
IMU and the orientation of the vehicle are considered to be the same at all times.
• It is assumed that when the inertial navigation algorithm begins, the initial linear velocity of the
vehicle is always zero (a.k.a. the vehicle starts at a standstill), the initial position of the vehicle is
always at the origin of the Navigation frame (0, 0, 0) and the initial heading of the vehicle is
always zero degrees (a.k.a. aligned with the X-axis of the Navigation frame). It is also assumed
that the heading of the vehicle is referenced to the local environment as opposed to the Earth’s
magnetic North Pole. Because the heading of the vehicle is locally referenced, its initial heading
can be set to any desired starting value and a value of zero degrees is chosen for the sake of
simplicity.
• It is assumed that the inertial data recorded at a specific time : is constant over the length of a
single timestep ∆:, meaning that the acceleration and angular velocity of the vehicle is assumed
to be constant from time : until immediately before the next timestep : + ∆: when a new sample
of inertial data is read from the IMU.

Section 5.3 – Inertial Navigation Algorithm Overview

This subsection provides a high-level overview of the inertial navigation algorithm. A basic block
diagram of the inertial navigation algorithm can be seen in Figure 5.3. As Figure 5.3 shows, the inertial
navigation algorithm is broken up into three phases. The first phase of the inertial navigation algorithm is
the initialization phase. The initialization phase occurs before the vehicle begins to move. During the
initialization phase, the initial linear velocity, position and orientation of the vehicle are determined.
Assumptions are made about the initial linear velocity and position of the vehicle, but determining the
initial orientation of the vehicle is more difficult. The initial roll and pitch of the vehicle are determined
using a process called self-leveling, where measurements from the accelerometers are used to determine

51
 Figure 5.3: A high-level block diagram of the inertial navigation algorithm.

how the IMU (and by extension, the vehicle) is oriented with respect to gravity. In outdoor environments,
the initial heading of the vehicle can be determined by using an external sensor system (such as a digital
magnetic compass (DMC) or a three-axis magnetometer) to measure the global heading of the vehicle
with respect to the Earth’s magnetic North Pole. In indoor environments or other environments where

52
globally referenced heading information isn’t strictly necessary, the initial heading of the vehicle can be
measured with respect to other features or local landmarks within the environment.
After the initialization phase ends and the vehicle begins to move, the algorithm enters the
estimation phase. During the estimation phase, inertial data from the IMU is used to produce initial
estimates of the linear velocity, position and orientation of the vehicle as it moves through its
environment. The effects of the static biases, scale factors and misalignment errors are removed from the
inertial data as well as the effects of other external factors such as the angular velocity caused by the
rotation of the Earth. Then, the corrected angular velocity data is used to estimate the orientation of the
vehicle. Once the orientation of the vehicle has been determined, the corrected acceleration data is
converted from the Body frame into the Navigation frame, the effects of gravity are removed and the
corrected acceleration data is integrated over time using Runge-Kutta numerical integration to estimate
the linear velocity and position of the vehicle. The end result of the estimation phase is initial estimates of
the linear velocity, position and orientation of the vehicle.
After the estimation phase, the algorithm enters the correction phase. In the correction phase, an
EKF is used to combine the dynamics of the stochastic errors (i.e. measurement noise processes, drifting
biases, etc.) and secondary navigation information to produce correction terms for the initial linear
velocity, position and orientation estimates produced in the estimation phase. These correction terms are
applied to the initial estimates from the estimation phase to produce the final estimates of linear velocity,
position and orientation for a specific timestep. However, the EKF can only generate correction terms if
secondary navigation information is available. So if there is a timestep where no new secondary
navigation information is available, then the estimation phase of the algorithm will be executed and the
estimation-error covariances of the EKF will be propagated, but no correction terms will be produced by
the EKF. In these instances, the initial estimates from the estimation phase of the algorithm become the
final estimates of linear velocity, position and orientation for that timestep.
After the correction phase, the algorithm loops back to the estimation phase and the process
described above is repeated using a new set of inertial measurements from the IMU and any new
secondary navigation information. The estimation and correction phases of the algorithm are repeated
indefinitely in this manner until the vehicle is either powered off or another goal condition is reached.
The following subsections will provide a more detailed description of each of the three phases of
the inertial navigation algorithm.

53
Section 5.4 – The Initialization Phase

This subsection explores the initialization phase of the inertial navigation algorithm in more detail
and explains how the initial linear velocity, position and orientation of the vehicle are determined. A basic
block diagram of the initialization phase can be seen in Figure 5.3. At the beginning of the algorithm
(time :q ), the vehicle has some known initial linear velocity, position and orientation. As mentioned
previously in Section 5.2, it is assumed that the initial linear velocity of the vehicle is zero, the initial
position of the vehicle is at the origin of the Navigation frame (0, 0, 0) and the initial heading of the
vehicle is zero degrees.
Before the vehicle begins to move, there is a sixty second period of time where the vehicle
remains stationary. During this time, the inertial navigation algorithm performs a process called self-
leveling. Self-leveling is very similar to the Multi-Angle Experiment described earlier in Section 4.3 in
that it uses gravity as a source of known and constant acceleration to determine the initial roll and pitch of
the IMU, and by extension the vehicle. These initial measurements of roll and pitch are used along with
the initial heading of the vehicle to create a direction cosine matrix (DCM) that describes the initial three-
dimensional orientation of the vehicle. The self-leveling routine is performed using a recursive linear least
squares estimation process that is very similar to one used in the Multi-Angle Experiment to estimate the
combined scale factor and misalignment error matrices for the accelerometers ( ). Refer to Section 4.3.1
for more details.

Section 5.5 – The Estimation Phase

This subsection explores the estimation phase of the inertial navigation algorithm in more detail
and explains how raw inertial measurements from the IMU are used to produce initial estimates of the
linear velocity, position and orientation of the vehicle. A detailed block diagram of the estimation phase
of the inertial navigation algorithm is shown in Figure 5.4.

position and orientation of the vehicle will be at the future timestep : + ∆:. At the current timestep : the
The purpose of the estimation phase is to calculate an initial estimate of what the linear velocity,

vehicle has some linear velocity, position and orientation, and the future linear velocity, position and
orientation of the vehicle at time : + ∆: is estimated using inertial information read from the IMU at the
current timestep :. As stated previously in Section 5.2, the inertial data recorded at a specific time : is
assumed to be constant over the length of a single timestep ∆:.

54
 Figure 5.4: A detailed block diagram of the estimation phase of the inertial navigation algorithm.

The estimation process begins by reading a set of raw inertial measurements (three-dimensional
acceleration, three-dimensional angular velocity and temperature) from the IMU. Based on the
temperature measurement, the corresponding static biases and scale factors for the accelerometers and
gyroscopes are retrieved from a look-up table and used to calculate compensation terms as shown in
Equations 5.1 and 5.2.

U= &+ Eq. 5.1

&= +?
U− +? Eq. 5.2

where & is the true acceleration or angular velocity, U is the measured value of the acceleration or angular
+? +?
velocity, is the scale factor and misalignment error compensation term and is the static bias
compensation term. These compensation terms are then applied to the raw inertial data to remove the

angular velocity terms are denoted by the variables %r 5 and s

effects of the static biases, scale factors and misalignment errors. The compensated acceleration and
t5D
5
.

55
 tuD
Next, another compensation term (s 5
) is applied to the angular velocity data to remove the
effects of the rotation of the Earth. The rotation of the Earth creates a very small angular velocity that is
constantly measured by the gyroscopes. Over time this angular velocity will propagate through the inertial
navigation algorithm and cause the estimated orientation of the vehicle to drift. The angular velocity of
the Earth is originally referenced to the Z-axis of the Earth frame, but it can be transformed into the
Navigation frame if the latitude, longitude and altitude of the vehicle are known. And because the vehicle
is assumed to be navigating on the LTP (see Section 5.2), the origin of the Navigation frame remains
fixed at a single point on the Earth’s surface and therefore the transformation between the Earth frame
and the Navigation frame always remains the same. This means that the angular velocity of the Earth can
be expressed in the Navigation frame as a constant three-dimensional angular velocity term. The quantity
s
tuD
5
tuD
is calculated by transforming the angular velocity of the Earth from the Navigation frame (s v
) into
the Body frame using the transpose of the DCM that represents the current orientation of the vehicle at
X
time :, which is wIx5v : y . The quantity s
tuD
5
is then subtracted from the compensated angular velocity

t5D
data (s 5
) to produce the quantity s
t5u
5
.

timestep : + ∆:. This is done by propagating the Ix5v : matrix forward through time based on the fully
The next step in the process is to calculate what the orientation of the vehicle will be at the future

compensated angular velocity data s

t5u
5
. Equations 5.3 through 5.6 define how a generic DCM is
propagated forward through time [16,17].

1 −∆z ∆z
Ix5v : + ∆: = Ix5v : ∆z 1 −∆z
−∆z ∆z 1
Eq. 5.3

∆z = s ∆: ∆z = s ∆: ∆z = s ∆: Eq. 5.4 – 5.6

The Ix5v : + ∆: matrix is then used to calculate the initial estimates of the roll, pitch and heading of the
vehicle at the future timestep : + ∆:.

data. First, the compensated acceleration data (%r 5 ) is converted from the Body frame into the Navigation
Now that the angular velocity data has been processed, the next step is to process the acceleration

frame using the DCM that represents the current orientation of the vehicle (Ix5v : ) and the effects of
gravity are removed as shown in Equation 5.7. This results in the fully compensated acceleration term
>Wuv : . This fully compensated acceleration term is then integrated over time using Runge-Kutta

56
the future timestep : + ∆: as shown in Equations 5.8 and 5.9 [18].
numerical integration to calculate the initial estimates of the linear velocity and position of the vehicle at

>Wuv : = Ix5v : %r 5 + {v Eq. 5.7

∆: v ∆:
NW : + ∆: = NW : + |>Wu : + 4>Wuv }: + ~ + >Wuv : + ∆: •
6 2
Eq. 5.8

∆: # v ∆:
€̂ : + ∆: = €̂ : + NW : ∆: + |>Wu : + 2>Wuv }: + ~•
6 2
Eq. 5.9

The initial estimates of linear velocity, position and orientation produced in this phase of the
inertial navigation algorithm accurately model the true linear velocity, position and orientation of the
vehicle over short periods of time, but imperfections and drifts in the IMU data will eventually cause the
estimates to diverge from the true linear velocity, position and orientation of the vehicle. In the next phase
of the algorithm, the correction phase, the dynamics of the stochastic IMU errors and secondary
navigation information will be incorporated into an EKF to produce correction terms for these initial
estimates. The correction terms generated by the EKF will help to correct the errors and drifts that build
up in the initial estimates of linear velocity, position and orientation due to the imperfections and drifts in
the inertial data returned by the IMU.

Section 5.6 – The Correction Phase

This subsection explores the correction phase of the inertial navigation algorithm in more detail
and explains how the dynamics of the stochastic IMU errors and secondary navigation information are
combined in an EKF to produce correction terms for the initial estimates of linear velocity, position and
orientation from the estimation phase of the algorithm. A block diagram of the correction phase of the
inertial navigation algorithm is shown in Figure 5.5.
The purpose of the correction phase is to generate correction terms for the initial estimates of
linear velocity, position and orientation that are produced in the estimation phase of the inertial navigation
algorithm. Specifically, there are fifteen correction terms that need to be generated: three each for the
accelerometer and gyroscope biases (denoted as ∆ and ∆ respectively), three for the linear velocity
estimates (denoted as ∆Nuv ), three for the position estimates (denoted as ∆€uv ) and three for the orientation
estimates (denoted as ∆z5v
5
). These correction terms are generated using an EKF. An EKF is a recursive

57
 Figure 5.5: A detailed block diagram of the correction phase of the inertial navigation algorithm.

Bayesian estimator that estimates one or more dynamic states over the course of time. The fifteen-element
state vector of the EKF is shown in Equation 5.10.

∆Nuv ∆€uv ∆ ∆z5v

X
& : = ‚∆ 5
ƒ Eq. 5.10

The basic operation of a generic EKF is as follows. An EKF is implemented with three main
steps: an initialization step, a propagation step and an update step. During the initialization step, an initial
estimate of the state vector (&W E 0 ) and an initial estimation-error covariance matrix („E 0 ) are
determined based on any knowledge that is available about the initial state of the system. During the
propagation step an a priori estimate of the state vector is generated based on the system model that
defines how the states behave and the estimation-error covariance matrix is updated based on the system
model as well. During the update step an a posteriori estimate of the state vector is generated based on the
a priori state estimate from the propagation step and either direct or indirect measurements of some or all
of the states in the state vector. The estimation-error covariance matrix is also updated based on the
measurement model that defines the relationship between the measurement data and the state vector. The
propagation and update steps are then repeated during each subsequent iteration of the EKF to produce
new estimates of the dynamic state vector over the course of time.

58
initialization step occurs the same way, and an initial estimate of the state vector (&W E 0 ) and an initial
The EKF presented in this thesis follows a slightly different approach than a generic EKF. The

estimation-error covariance matrix („E 0 ) are determined based on any knowledge that is available
about the initial state of the system (as shown in Equations 5.11 and 5.12).

&W E 0 = 80 0 0 0 09X Eq. 5.11

"# 0 0 0 0
_ 5 c
^ 0 0 0 0b
„E 0 =^ 0 0 0 0b
^ 0 0 0 "5# 0b
Eq. 5.12

] 0 0 0 0 a

The initial estimate of the state vector (&W E 0 ) is set to zero because during the first timestep the
linear velocity, position and orientation of the vehicle are already known so no correction terms are
necessary. The initial estimation-error covariance matrix („E 0 ) is set up as a diagonal matrix with
Identity matrices along its diagonal except for the positions that correspond to the bias correction states.
These positions are populated with the turn-on to turn-on bias variation terms calculated in Section 4.2
because they represent the initial variability in the values of the static biases.
After the initialization step, the EKF moves on to the propagation step. During the propagation
step, the dynamics of the stochastic IMU errors are used to generate the a priori estimation-error
covariance matrix, but no a priori state estimates are generated. The a priori estimate of the state vector is
simply set to contain all zeroes.
The equations for implementing the propagation step of the EKF are shown in Equations 5.13
through 5.16 [19,20].

−1 0 0 0 0
_ xv c
^−I5 : 0 0 0 I5 : H−%r 5 : ×Jb
x v

G : =^ 0 0 0 0 b
^ 0 0 0 −1 0 b
^ b
Eq. 5.13

] 0 0 0 − tuD
H−s : ×J a
5

"5/ #
0 0 0 0
_ c
^ 0 "# 0 0 0 b
†=^ 0 0 "‡# 0 0 b
^ b
0 0 0 0 b
#
H"5/ J
Eq. 5.14

^
] 0 0 0 0 "# a

ˆ : = + G : ∆: Eq. 5.15

59
 „+ : = ˆ : „E : − ∆: ˆ : X
+ †∆: Eq. 5.16

The terms in the state transition matrix (G : ) show how each of the states in the state vector
(& : ) are related to each other at the current timestep. The variance terms along the diagonal of the
process noise covariance matrix (†) represent the inherent randomness of each of the states in the state
vector (& : ). The first and fourth variance terms along the diagonal represents the variances of the WGN
processes that drive the drifting biases of the accelerometers and gyroscopes respectively. The second and
fifth variance terms along the diagonal represents the variances of the measurement noises for the
accelerometers and gyroscopes respectively. Finally, the third variance term along the diagonal represents
the variances of the position correction states in the state vector (∆€uv ).

timestep („+ : ) is calculated based on the state transition matrix for the current timestep (G : ), the a
Equations 5.15 and 5.16 define how the a priori estimation-error covariance matrix for the current

posteriori estimation-error covariance matrix from the previous timestep („E : − ∆: ) and the process
noise covariance matrix (†).
After the propagation step, the EKF moves on to the update step. During the update step, the EKF
uses the a priori estimation-error covariance matrix from the propagation step and any available
secondary navigation information to generate a set of correction terms for a specific timestep. As
mentioned previously, the update step of the EKF is only executed when new secondary navigation
information is available. If there is no new secondary navigation information for a specific timestep, then
correction terms cannot be generated and only the propagation step of the EKF is executed. Typically, this
secondary navigation information is produced by a separate sensor system such as GPS or a visual
odometry system. The secondary navigation information for the inertial navigation algorithm presented in
this thesis is generated using a combination of zero velocity updates (ZUPTs), position fixes and
orientation fixes. A ZUPT is a period of time when the vehicle is known to have both a linear and angular
velocity of zero. If the ZUPT occurs at a known location, then the position and orientation of the vehicle
can be recorded as well. These accurate measurements of the true position and orientation of the vehicle
are referred to as position and orientation fixes respectively.
The equations for implementing the update step of the EKF are shown in Equations 5.17 through
5.22 [19,20].

∆€uv ∆ ∆z5v
X
U : = ‚∆Nuv 5
ƒ Eq. 5.17

60
 0 0 0 0
0 0 0 0
V=
0 0 0 0
Eq. 5.18

0 0 0 0

"# 0 0 0
_ ∆‰ c
^ 0 "∆‡ 0 0 b
#
I=^
^ 0 0 "∆5
#
0 bb
Eq. 5.19

] 0 0 0 "∆Š
#
a

M : = „+ : V X V„+ : V X + I +? Eq. 5.20

&W E : = M : U : Eq. 5.21

„E : = − M : V „+ : −M : V X
+ M : IM : X Eq. 5.22

The terms in the measurement data vector (U : ) represent the twelve secondary navigation
information terms that are recorded when the vehicle undergoes a ZUPT at a specific time :. The terms in
the measurement matrix (V) show how the terms in the measurement data vector (U : ) are related to the
terms in the state vector (& : ). The variance terms along the diagonal of the measurement noise
covariance matrix (I) represent the variances of each of the measurements in the measurement data
vector (U : ). The first variance term along the diagonal represents the variances of the measurements
that are made of the velocity correction states (∆Nuv ). The second term along the diagonal represents the
variances of the measurements that are made of the position correction states (∆€uv ). The third term along
the diagonal represents the variances of the measurements that are made of the gyroscope bias correction
states (∆ ). And finally, the fourth term along the diagonal represents the variances of the measurements
5
that are made of the orientation correction states (∆z5v ).
Equation 5.20 defines how the optimal gain matrix for the current timestep (M : ) is calculated
based on the a priori estimation-error covariance matrix for the current timestep („+ : ), the
measurement matrix (V) and the measurement noise covariance matrix (I). Equation 5.21 defines how
the a posteriori estimate of the state vector for the current timestep (&W E : ) is calculated based on the
optimal gain matrix for the current timestep (M : ) and the measurement data vector for the current
timestep (U : ). And finally, Equation 5.22 defines how the a posteriori estimation-error covariance
matrix for the current timestep („E : ) is calculated based on the a priori estimation-error covariance
matrix for the current timestep („+ : ), the optimal gain matrix for the current timestep (M : ), the
measurement matrix (V) and the measurement noise covariance matrix (I).

61
 After the update step, assuming that secondary navigation information was available during the
current timestep and the update step was successfully performed, the fifteen correction terms generated by
the EKF are used to update the quantities that they correspond to within the inertial navigation algorithm.
The equations for applying the correction terms generated by the EKF are shown in Equations
5.23 through 5.27 [20].

x = x +∆ Eq. 5.23

NWuv = NWuv + ∆Nuv Eq. 5.24

€̂uv = €̂uv + ∆€uv Eq. 5.25

x = x +∆ Eq. 5.26

zxuv = zxuv + ∆z5v

5 Eq. 5.27

Once the fifteen quantities in the above equations have been updated, the correction phase of the
inertial navigation algorithm is over. The inertial navigation algorithm moves on to the next timestep and
the estimation and correction phases are repeated using a new set of inertial data from the IMU and a new
set of secondary navigation information (if any is available). The estimation and correction phases of the
inertial navigation algorithm are repeated indefinitely in this manner until the vehicle is either powered
off or another goal condition is reached.

62
Section 6 – UGV Navigation Test Plan

Once the inertial navigation algorithm has been completely developed and coded, it must be
tested using inertial data from a real-world navigation scenario. To acquire this data, both the SparkFun
and ADI IMUs were rigidly mounted to a UGV, which was then driven through a predefined test course.
The UGV was driven through the test course three times (starting and ending in the same place each
time), resulting in three independent sets of inertial data for each IMU. The inertial data recorded from
both IMUs during these tests was used to determine how accurately the inertial navigation algorithm can
track the linear velocity, position and orientation of a moving vehicle. This section will explain the overall
methodology behind the UGV testing, including how inertial data was logged from the two IMUs, how
the test course was set up, how the vehicle was driven and how the true linear velocity, position and
orientation of the vehicle was determined.

Section 6.1 – The Test Course

The UGV test course was set up on the third floor of the Atwater Kent building on the Worcester
Polytechnic Institute campus. This location was chosen for several reasons. Indoor environments (and
hallways in particular) are typically very structured and flat with few obstructions. This means that any
vehicle driving through an indoor environment is essentially navigating along a single planar surface that
is also level with gravity. Deviations from this ideal surface (i.e. uneven or un-level patches of the floor)
are expected to be fairly minimal and shouldn’t have much of an effect on the vehicle’s ability to
maneuver through the environment. Indoor environments are also simple to measure and map, which is
essential for accurately placing position markers to gather truth data for the inertial navigation algorithm.
The test course itself is rectangular in shape and measures 22 meters long by 16.5 meters wide. A
rectangular test course was chosen so that the vehicle could travel in a loop and finish at the same location
that it started at. This loop closure is a useful metric to have because after one complete loop of the test
course, the vehicle should ideally be at the exact same location that it started at. Any discrepancies
between the estimated starting and finishing location of the vehicle can be used as a metric to assess the
accuracy of the inertial navigation algorithm.
To generate secondary navigation information for the EKF component of the inertial navigation
algorithm (as described in Section 5.6), the vehicle must be periodically stopped at specific known
locations along the test course so that ZUPTs, position fixes and orientation fixes can be recorded. In
order to mark the locations where the vehicle should be stopped, position markers were laid out at two

63
 Figure 6.1: The position markers outlining the locations of ZUPTs along the length of the test course.

meter intervals along the entire length of the test course as seen in Figure 6.1 (with the exception of two
position markers that had to be laid out at two and a half meter intervals due to the 16.5 meter length of
two of the hallways). The position markers were laid out directly in the center of the hallways, and also
served as convenient waypoints to help keep the vehicle from straying off of the predefined test course.
Since the position markers were laid out at known intervals, the position and orientation of the vehicle can
be accurately determined when it is stopped on top of a specific marker.
In a more elaborate INS, the secondary navigation information would come from a physical
sensor system (GPS for example) at a frequency on the order of 1Hz or faster. However, due to the
spacing of the position markers and the relatively low top speed of the vehicle, secondary navigation
information was only recorded at approximately 0.1Hz (approximately once every ten seconds). Because
of the long stretches in between secondary navigation information readings, the drifts and other
imperfections in the inertial data have a longer amount of time to accumulate.

64
Section 6.2 – Configuring and Driving the UGV

The vehicle used to navigate the test course was a Husky A100 UGV from the company
Clearpath Robotics. A picture of the Husky A100 UGV can be seen in Figure 6.2. The Husky A100 UGV
is a six-wheeled, skid-steered vehicle, which makes it ideal for outdoor environments with rough terrain.
However, this wheel configuration can present some problems when trying to maneuver on the carpeted
surfaces typically found in many indoor environments. Because of the high-traction, all-terrain treads of
the vehicle’s wheels and the skid-steering nature of how the vehicle turns there is a lot of friction that
builds up between the wheels of the vehicle and the carpet. This friction results in turning movements that
are jerky and non-fluid. To get around this issue, each of the vehicle’s six wheels was coated in a single
layer of duct tape (as seen in Figure 6.3) which significantly reduced the traction of the vehicle’s tires as
well as the coefficient of friction between the vehicle’s wheels and the carpet. This reduction in friction
allowed the vehicle to make turns that were much more fluid and continuous.
Before each test run, both the SparkFun and ADI IMUs were rigidly mounted to the vehicle as
shown in Figure 6.4. At the beginning of each test run, the vehicle was kept stationary for sixty seconds
while the two IMUs recorded inertial data. The data from this initial stationary period was used to

Figure 6.2: The Husky A100 UGV.

65
Figure 6.3: The wheels of the Husky A100 UGV coated in duct tape to reduce the coefficient of friction between the wheels
and the carpeted floor.

Figure 6.4: The SparkFun and ADI IMUs rigidly mounted to the Husky A100 UGV.

66
 Figure 6.5: The Husky A100 UGV positioned at the beginning of the test course.

perform the self-leveling routine discussed in Section 5.4. The vehicle initially began each test run at one
of the corners of the test course as shown in Figure 6.5. It was then manually driven through a single
complete loop of the test course, stopping for ten seconds at each one of the designated position markers
to record ZUPTs, position fix and orientation fix data, until it arrived back at its original starting location.
Throughout the test run, inertial data was continuously logged from both the SparkFun and ADI IMUs at
a rate of 25Hz.
Another feature of the Husky A100 UGV is that it has a built in speed control algorithm, which
means that it can be commanded to drive forwards, backwards or turn in either direction at a specific rate.
The forward velocity of the vehicle was kept at a constant rate of 0.5 meters per second whenever
possible during the course of a test run. The vehicle only accelerated or decelerated when it was either
starting from or slowing to a stop respectively. Likewise, the turning velocity of the vehicle was also kept
at a constant rate of approximately 0.4 rad/s throughout the course of any turning maneuvers, and it would
only accelerate or decelerate when it was either just starting or just ending a turning maneuver
respectively.

67
Section 7 – Results and Discussion

This section presents and discusses the results obtained from the UGV navigation testing
described in Section 6. All of the results presented in this section are from the same navigation testing
trial (Trial #3). The complete results from all three of the navigation testing trials are located in the
Appendix. The first three subsections contain the results obtained from running the two INS units through
some specific navigation scenarios and the corresponding discussion of those results. The final subsection
contains a direct comparison of the two INS units and an evaluation of which INS was able to produce the
most accurate estimates of the linear velocity, position and orientation of the UGV over the length of the
test course.

Section 7.1 – Inertial Data Only

The results presented in this subsection illustrate how the SparkFun-based and ADI-based INS
units perform when they are limited to using only inertial data from the IMUs and no secondary
navigation information. Figure 7.1 shows the estimates of the position and orientation of the vehicle over
time produced by the two INS units versus the true position and orientation of the vehicle over time.
Figure 7.1 shows just how important secondary navigation information is to the success of the
inertial navigation algorithm. With no secondary navigation information to help compensate for the

Figure 7.1: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (Inertial data
only).

68
inaccuracies and drifts in the inertial data, the errors in the estimated linear velocity, position and
orientation of the vehicle grow unbounded over time. If left unchecked these errors become very large
very quickly, which is why accurately compensating for the drifts and other imperfections in the inertial
data using secondary navigation information is a huge part of successfully navigating using inertial
sensors.

Section 7.2 – Inertial Data and ZUPTs

The results presented in this subsection illustrate how the SparkFun-based and ADI-based INS
units perform when they are limited to using only inertial data from the IMUs and velocity fixes (i.e.
ZUPTs). Figure 7.2 shows the estimates of the position and orientation of the vehicle over time produced
by the two INS units versus the true position and orientation of the vehicle over time.
Compared to the results from Section 7.1, the results in Figure 7.2 clearly show that the linear
velocity information obtained from periodically performing ZUPTs helps to significantly limit the errors
in the estimated linear velocity, position and orientation of the vehicle over time. The secondary
navigation information obtained from these ZUPTs allows the EKF to generate correction terms for the
initial linear velocity estimates based on the inertial data from the IMU. Applying these correction terms
to the initial linear velocity estimates produces final estimates of the linear velocity, position and
orientation of the vehicle that are much more accurate.

Figure 7.2: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (Inertial data
and ZUPTs only).

69
Section 7.3 – Inertial Data, ZUPTs and Position and Orientation Fixes

The results presented in this subsection illustrate how the SparkFun-based and ADI-based INS
units perform when they are free to use all of the information that is available to them (i.e. inertial data
from the IMUs, velocity fixes (ZUPTs) and position and orientation fixes). These results represent the
navigation capabilities of the fully completed INS units. Figure 7.3 shows the estimates of the position
and orientation of the vehicle over time produced by the two INS units versus the true position and
orientation of the vehicle over time.
As shown previously in Section 7.2, the addition of secondary linear velocity information into the
system greatly increased the accuracy of the estimates produced by both INS units. Now, by adding
secondary position and orientation information into the system as well, the accuracy of the estimates is
increased even further. Figure 7.3 shows how an inertial navigation system can produce accurate linear
velocity, position and orientation estimates when it is provided with enough secondary navigation
information to sufficiently compensate for the drifts and other imperfections in the IMU data. The
estimates are still not perfect, but the secondary navigation information for these experiments was
collected at a relatively low frequency (approximately once every ten seconds). In most commercial, real-
time INS units, secondary navigation information is gathered much more quickly (approximately once per
second or faster), so EKF corrections are typically generated at a much higher frequency. And as EKF
corrections are generated more and more frequently, the estimates produced by the INS become more and
more accurate.

Figure 7.3: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (Inertial data,
ZUPTs and position and orientation fixes).

70
Section 7.4 – SparkFun and ADI IMU Comparison

The results presented in this final subsection provide a direct comparison of the navigation results
generated by the SparkFun-based INS and the ADI-based INS. For the purposes of this comparison, each
INS had access to the inertial data from its respective IMU and all of the available secondary navigation
information collected during the UGV navigation tests (i.e. ZUPTs, position fixes and orientation fixes).
The two INS units are compared in two ways. The first comparison is a visual comparison of how
well each INS was able to track the linear velocity, position and orientation of the vehicle over time as it
drove around the predefined test course. The second comparison is a numeric comparison of the
estimation errors from each of the INS units (i.e. the differences between the estimated linear velocity,
position and orientation from the INS units and the true linear velocity, position and orientation of the
vehicle over time). These comparisons are shown in Figures 7.4 through 7.7. Figure 7.4 displays the
estimated position and orientation of the vehicle over time from each of the INS units on the same graph
versus the true position and orientation of the vehicle over time. Figure 7.5 displays the estimated linear
velocity of the vehicle over time from each of the INS units on the same graph versus the target velocities
for the vehicle over time. And finally, Figures 7.6 and 7.7 display the errors between the linear velocity,
position and orientation estimates produced by each of the INS units and the true linear velocity, position
and orientation of the vehicle over time.

Figure 7.4: A visual comparison of the position and orientation estimates produced by the SparkFun-based and ADI-based
INS units.

71
 Figure 7.5: A visual comparison of the linear velocity estimates produced by the SparkFun-based and ADI-based INS units.

Figure 7.6: Errors in the estimated linear velocity, position and orientation of the vehicle over time (ADI).

As Figures 7.4 through 7.7 show, the navigation data from the ADI INS is clearly much more
accurate than the navigation data from the SparkFun INS is in terms of both the visual comparison and the
numerical comparison. This result is intuitively satisfying given the cost differential between the two
IMUs, and it supports the common-sense hypothesis that as the cost per unit of a MEMS IMU increases,
the quality of the IMU and of the inertial data that it returns increases proportionally.

72
 Figure 7.7: Errors in the estimated linear velocity, position and orientation of the vehicle over time (SparkFun).

The primary conclusion to take away from these results is that the analysis and modeling of
typical MEMS IMU error sources did make the inertial data from the SparkFun IMU better overall, and
the performance gap between the SparkFun and ADI IMUs did decrease, but the inertial data from the
SparkFun IMU is still not as high-quality as the inertial data from the ADI IMU is. The inertial data from
the ADI IMU is still much better overall, due largely in part to how the internal mechanisms of the two
IMUs are set up. While both the SparkFun and ADI IMUs are constructed using MEMS technology, the
ADI IMU has several sophisticated internal mechanisms (temperature calibration, low-pass filtering, etc.)
that operate on the inertial data before it is received by the user. By comparison, the SparkFun IMU is
much simpler and contains less sophisticated internal mechanisms for preprocessing the raw data from its
inertial sensors. Since the inertial data from the ADI IMU is already cleaned up to some extent, it is
inherently more accurate than the inertial data from the SparkFun IMU is. Because of this, the ADI IMU
has a distinct advantage over the SparkFun IMU in terms of producing accurate navigation information.
Furthermore, because each of the INS units uses the exact same secondary navigation information
to calculate their EKF correction terms, any differences in the final navigation results are entirely related
to the quality of the inertial data produced by the IMUs. Therefore, the suitability of each IMU for
navigation applications can be assessed by observing which INS produced the better overall navigation
results. And given the results displayed in Figures 7.4 through 7.7, it can be concluded that the ADI IMU
definitely produces better overall inertial data than the SparkFun IMU does, and is therefore much better
suited for inertial navigation applications than the SparkFun IMU is.

73
Section 8 – Future Work

This section lists several new ideas that could potentially be implemented in future iterations of
this research to either improve the quality of the results obtained from the IMU calibration experiments or
the inertial navigation algorithm or add a new component or functionality to the existing research that has
already been accomplished.

• Perform unit to unit testing. Testing multiple different IMUs of the same model number would
help to determine how widely errors vary in between different units. Unit to unit testing would
determine whether or not each individual IMU must be separately analyzed or if the results from
a single analysis would be enough to model the errors for all IMUs of a specific model number.
• Automate the IMU laboratory experiments. Many of the experiments used to estimate the IMU
error parameters require the same action to be performed over and over again. These experiments
could be greatly improved and simplified by automating the process.
• Analyze the performance of the EKF in more depth. There are many aspects of Kalman Filtering
not touched upon in this thesis, and they could be explored in more detail to further improve the
performance of the EKF implemented in the inertial navigation algorithm. Some examples
include analyzing the estimation-error covariance matrices („‹ ) and analyzing the EKF
innovations (U‹ − V&W‹+ ).
• Find a new testing location and/or testing vehicle. Using a six-wheeled, skid-steered vehicle on a
high-friction carpeted surface doesn’t make for the best testing environment. In future iterations
of this research or similar research, it would make sense to find a vehicle with a smoother steering
mechanism as well as a testing location with a lower coefficient of friction. This will hopefully
result in smoother vehicle motions during the duration of navigation testing.
• Improve the test course. The test course that was used for the navigation tests performed in this
thesis proved to be adequate. However, in order to test navigation performance with a higher
degree of accuracy, a more structured test course with a higher degree of overall precision would
be required. An improved method for gathering secondary navigation information for the inertial
navigation algorithm would also be required.
• Use real secondary navigation information. Secondary navigation information could be obtained
much more frequently from a real sensor system (GPS, stereo camera systems, etc.). More
frequent secondary navigation information would lead to more frequent EKF correction terms,
which in turn would greatly increase the accuracy of the navigation testing results.

74
Section 9 – Conclusion

The goal of this thesis was to design, implement and test two separate INS units based around two
low-cost MEMS IMUs. The results of the testing procedures were then used to determine how well each
individual INS performed and which IMU is better suited for implementation in an INS. To achieve this
goal, each IMU was first analyzed and modeled using a series of laboratory experiments designed to
isolate and characterize the behavior of several well-documented types of MEMS IMU errors. Then, an
inertial navigation algorithm was designed that incorporated the information and error models produced
by the laboratory experiments. The inertial navigation algorithm was tested by rigidly mounting the two
IMUs to a UGV and then driving the UGV via remote control through a predefined test course. The
inertial information recorded from each IMU during the navigation tests was then post processed using
the inertial navigation algorithm. Finally, the resulting navigation information was analyzed and
compared to the true linear velocity, position and orientation of the UGV to determine which INS was
able to more accurately track the linear velocity, position and orientation of the UGV over time.
There are several primary results to take away from this thesis. The first result is that the analysis
and modeling of typical MEMS IMU error sources did make the inertial data from the SparkFun IMU
better overall, but it was not enough to bring the SparkFun IMU up to the performance level of the ADI
IMU. The ADI IMU proved to ultimately be much better for inertial navigation applications than the
SparkFun IMU was. The second result is that while inertial data is the primary component of an INS,
secondary navigation information plays an equally important role in the overall success of the INS. As
secondary navigation information is obtained more frequently, the results from the INS will improve
proportionally. The third and final result is that if high quality inertial data is essential to the success of a
system, the best way to obtain that data is most likely to invest in a high quality IMU with internal
mechanisms that preprocess the raw data from the inertial sensors. Low-cost, hobbyist-grade IMUs such
as the SparkFun IMU are better suited for applications where high quality inertial data is not essential,
such as impulse recognition, sensor fusion with a second type of sensor, wheel slippage detection,
unintended vehicle motion detection and self-leveling or orientation detection.
Overall, this thesis showed that it is possible to analyze and model several types of common
MEMS IMU errors and design an INS that uses those error models along with inertial data and secondary
navigation information to estimate the linear velocity, position and orientation of a UGV over time to
within a certain degree of accuracy. However, in order for the INS to be successful, the IMU supplying
the inertial data to the system must be high quality enough to produce accurate measurements of the
accelerations and angular velocities that the vehicle experiences over time.

75
Appendix

The Appendix displays all of the results generated by both the SparkFun-based and ADI-based
INS units for all three of the navigation testing trials. Both of the INS units were allowed to use the
inertial data from their respective IMUs and all of the secondary navigation information that was
available to them (i.e. ZUPTs, position fixes and orientation fixes). Each of the following subsections
displays the full results from a specific UGV navigation testing trial. The complete results include graphs
displaying the following information:

• The estimated linear velocity of the vehicle versus the desired target velocities for the vehicle
• The estimated position and orientation of the vehicle versus the true position and orientation of
the vehicle
• The errors between the linear velocity, position and orientation estimates produced by the two
INS units and the true linear velocity, position and orientation of the vehicle over time

Section A.1 – Full Results of UGV Navigation Testing (Trial #1)

Figure A.1: Estimated linear velocity of the vehicle versus target velocities (ADI).

76
 Figure A.2: Estimated linear velocity of the vehicle versus target velocities (SparkFun).

Figure A.3: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (ADI).

77
Figure A.4: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (SparkFun).

Figure A.5: Linear velocity, position and orientation estimation errors over time (ADI).

78
Figure A.6: Linear velocity, position and orientation estimation errors over time (SparkFun).

79
Section A.2 – Full Results of UGV Navigation Testing (Trial #2)

Figure A.7: Estimated linear velocity of the vehicle versus target velocities (ADI).

Figure A.8: Estimated linear velocity of the vehicle versus target velocities (SparkFun).

80
 Figure A.9: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (ADI).

Figure A.10: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (SparkFun).

81
 Figure A.11: Linear velocity, position and orientation estimation errors over time (ADI).

Figure A.12: Linear velocity, position and orientation estimation errors over time (SparkFun).

82
Section A.3 – Full Results of UGV Navigation Testing (Trial #3)

Figure A.13: Estimated linear velocity of the vehicle versus target velocities (ADI).

Figure A.14: Estimated linear velocity of the vehicle versus target velocities (SparkFun).

83
 Figure A.15: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (ADI).

Figure A.16: Estimated position and orientation of the vehicle versus true position and orientation of the vehicle (SparkFun).

84
 Figure A.17: Linear velocity, position and orientation estimation errors over time (ADI).

Figure A.18: Linear velocity, position and orientation estimation errors over time (SparkFun).

85
References

[1] Analog Devices, Inc., “3-Axis, ±2g/±4g/±8g/±16g Digital Accelerometer,” ADXL345 datasheet,
2009.
[2] InvenSense, Inc., “ITG-3200 Product Specification Revision 1.4,” ITG-3200 datasheet, March 2010.
[3] Analog Devices, Inc., “Six Degrees of Freedom Inertial Sensor,” ADIS16364 datasheet, 2009-2011.
[4] J. C. Hung, J. R. Thacher, H. V. White, “Calibration of Accelerometer Triad of an IMU With Drifting
Z-Accelerometer Bias,” In Proc. IEEE 1989 National Aerospace and Electronics Conference, 1989, pp.
153-158 Vol. 1.
[5] L. Zhang et al., “Research on Auto Compensation Technique of Strap-Down Inertial Navigation
Systems,” International Asia Conference on Informatics in Control, Automation and Robotics, 2009, pp.
350-353.
[6] B. Hartmann, N. Link, G. F. Trommer, “Indoor 3D Position Estimation Using Low-Cost Inertial
Sensors and Marker-Based Video-Tracking,” 2010 IEEE/ION Position Location and Navigation
Symposium (PLANS), 2010, pp. 319-326.
[7] L. Wang et al., “Thermal Calibration of MEMS Inertial Sensors for an FPGA-Based Navigation
System,” 3rd International Conference on Intelligent Networks and Intelligent Systems (ICINIS), 2010,
pp. 139-143.
[8] D. Unsal, K. Demirbas, “Estimation of Deterministic and Stochastic IMU Error Parameters,” 2012
IEEE/ION Position Location and Navigation Symposium (PLANS), 2012, pp. 862-868.
[9] J. J. Higgins, S. Keller-McNulty, Concepts in Probability and Stochastic Modeling. Boston, MA,
1995, pg. 330, Theorem 8.2-1.
[10] D. W. Allan, “Time and Frequency (Time-Domain) Characterization, Estimation, and Prediction of
Precision Clocks and Oscillators,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency
Control, vol. 34, no. 6, pp. 647-654, November, 1987. [Online serial]. Available:
http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=1539968&url=http%3A%2F%2Fieeexplore.ieee.o
rg%2Fxpls%2Fabs_all.jsp%3Farnumber%3D1539968. [Accessed 2011].
[11] O. J. Woodman, “An Introduction to Inertial Navigation,” University of Cambridge Computer
Laboratory, Cambridge, United Kingdom, Tech. Report. UCAM-CL-TR-696, Aug. 2007.
[12] P. D. Welch, “The Use of Fast Fourier Transform for the Estimation of Power Spectra: A Method
Based on Time Averaging Over Short, Modified Periodograms,” IEEE Transactions on Audio and
Electroacoustics, vol. 15, no. 2, pp. 70-73, June, 1967. [Online serial]. Available:

86
http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=1161901&url=http%3A%2F%2Fieeexplore.ieee.o
rg%2Fxpls%2Fabs_all.jsp%3Farnumber%3D1161901. [Accessed 2012].
[13] S. Vitale et al., “Recommendation for an algorithm for Power Spectral Density estimation for LISA
Pathfinder,” Università Degli Studi Di Trento, Trento, Italy, Tech. Report. S2-UTN-TN-3040, Sept. 2006.
[Online]. Available:
http://www.rssd.esa.int/SP/LISAPATHFINDER/docs/Data_Analysis/PSD_Recipe.pdf. [Accessed 2012].
[14] Penn State. STAT 510. Class Lecture, Topic: “Applied Time Series Analysis.” Department of
Statistics, Eberly College of Science, Pennsylvania State University, University Park, State College, PA,
2013. [Online]. Available: https://onlinecourses.science.psu.edu/stat510/?q=node/60. [Accessed 2013].
[15] S. M. Kay, “Chapter 4 – Linear Models,” in Fundamentals of Statistical Signal Processing (Volume
I: Estimation Theory). Upper Saddle River, NJ, 1993, pp-83-100.
[16] D. Titterton, J. L. Weston, “Strapdown Attitude Representations,” in Strapdown Inertial Navigation
Technology, 2nd Edition. IET, 2004, pp-36-44.
[17] W. Premerlani, P. Bizard, Direction Cosine Matrix IMU: Theory, May 2009.
[18] J. L. Center, Jr., Integration for Local Tangent Navigation, Autonomous Exploration, Inc., Aug.
2013.
[19] D. Simon, Optimal State Estimation. Hoboken, NJ, 2006.
[20] J. L. Center, Jr., Information Filter Signal Flow, Autonomous Exploration, Inc., 2011.

87