International Journal of Computer Applications Technology and Research

Volume 2– Issue 1, 32-36, 2013, ISSN: 2319–8656

Speed Determination of Moving Vehicles using Lucas-

Kanade Algorithm
Dolley Shukla Ekta Patel
Shri Shankaracharya College of Engg. & Tech. Shri Shankaracharya College of Engg. & Tech.
Junwani, Bhilai-490020 Junwani, Bhilai- 490020
Durg,(CG), India Durg,(CG), India

Abstract: This paper presents a novel velocity estimation method for ground vehicles. The task here is to automatically estimate
vehicle speed from video sequences acquired with a fixed mounted camera. The vehicle motion is detected and tracked along the
frames using Lucas-Kanade algorithm. The distance traveled by the vehicle is calculated using the movement of the centroid over the
frames and the speed of the vehicle is estimated. The average speed of cars is determined from various frames. The application is
developed using MATLAB and SIMULINK.

Keywords: Tracking, Optical flow, Motion estimation, Lucas-Kanade algorithm,velocity

1. INTRODUCTION identify moving objects, calculate object trajectory and to find

their velocity.
1.1 Video and Image Sequence:
Video is the technology of electronically capturing, In this paper, we present an efficient method for
recording, processing, storing, transmitting, and computing direction of motion of a vehicle and estimate its
reconstructing a sequence of still images representing scenes speed.[1] The proposed method firstly uses optical flow
in motion.[1] An image is a rectangular grid of pixels. It has a algorithm to calculate changes in the intensity of the pixels of
definite height and a definite width counted in pixels. the images. These apparent velocity components are then
A video usually consists of scenes, and each scene subjected to image processing techniques to obtain centroid of
includes one or more shots. A shot is an uninterrupted the vehicle across the frames. The image coordinates of the
segment of video frame sequence with static or continuous centroid are mapped to World space. Using this information
camera motion, while a scene is a series of consecutive shots the velocity of the vehicle is estimated.
that are coherent from the narrative point of view.

Video Sequence

Scenes

Fig2: Motion vector definition of current frame a) at

time t-1 b) at time t
Shot
1.3 Methods of Motion Estimation:

1.3.1 Feature/Region Matching: Motion is estimated

by correlating or matching features (e.g. edges) or regional
Frames and Key Frames(Red)
intensities (e.g. blocks of pixels) from one frame to another.
Fig1: Hierarchical structure of Video Sequence [1]
Examples include:
• Block-matching algorithm
1.2 Motion estimation:
• Phase correlation and frequency domain methods
Motion estimation is an important task of video
analysis. [1] It can be used to find the motion fields, to
1.3.2 Gradient based methods: Gradient-based
methods use spatial and temporal partial derivatives to
estimate image flow at every position in the image.

www.ijcat.com 32
 International Journal of Computer Applications Technology and Research
Volume 2– Issue 1, 32-36, 2013, ISSN: 2319–8656

• Optical flow relationship of a tracked vehicle across frames is determined

• Pixel recursive algorithms through transformation of pixel locations from image to world
coordinate system. Transformation to world coordinates and
pixel calibration are achieved by using standard lane markings
2. LITERATURE REVIEW whose length and gap distance are standard and known.
Mehrubeoglu and McLauchlan[2] detect and count
vehicles during day and night scenarios and different He et al. developed an embedded system to take traffic
environmental conditions in 2006. This paper is an extension measurements [9]. The authors used background subtraction
of the authors‟ work from detecting vehicles in still images to to aid in vehicle detection. The researchers then used
tracking vehicles in video. parallelograms for the regions of interest (ROI) due to the
image distortion resulting from the camera position and to
In their paper, An et al. report a motion tracking reduce the computational load.
algorithm that tracks different objects in motion in video.[3]
The authors achieve motion tracking by segmenting key
3. METHODOLOGY
frames, and then clustering objects whose motion is close to
the previously detected objects. Classification is achieved 3.1 Implementation in Simulink:
with a distance measure based on epicenter geometry.
The Simulink model for this project mainly consists of
three parts, which are “Velocity Estimation”, “Velocity
Yu et al. describe an algorithm that estimates traffic
Threshold Calculation” and “Object Boundary Box
density and average speed from Skycam MPEG compressed
Determination”.
images.[4] The authors compute DCT coefficients and
analyze motion vector projections across frames. Direction,
magnitude and texture filters are used to eliminate redundant
motion vectors. After mapping the image plane to world
coordinates, the average vehicle speed is estimated over a
video clip of 10 seconds, with a frame rate of 10 fps.

Anagnostopoulos et al. surveyed license plate

recognition methods.[5] The authors broke license plate
recognition into three parts, 1) license plate location, 2)
license plate segmentation, and 3) character recognition. In
addition, the authors have devised a database of license plate
images and videos under varying lighting and environmental
conditions that researchers may utilize as a common test set to
enable comparisons of various algorithms. Fig3: Simulink Block Diagram for Tracking Moving
Objects Using Lucas-Kanade Algorithm
Garibotto et al. utilize computer vision techniques to
estimate a vehicle‟s speed from two images by tracking 3.2 Lucas-Kanade Algorithm:
license plates and determining the distance traveled.[6] The
speed is calculated by using this traveled distance and the time The Lucas–Kanade method is a two-frame differential
difference between the two images. The researchers describe method for optical flow estimation developed by Bruce D.
both monocular as well as binocular vision systems. In our Lucas and Takeo Kanade.
method, license plate information is not needed.
It introduces an additional term to the optical flow by
Pelegrí et al. developed and tested GMR magnetic assuming the flow to be constant in a local neighbourhood
sensors to determine car speeds.[7] The vehicle causes around the central pixel under consideration at any given
changes in the magnetic field of the sensor when it travels time.[10]
over the sensor. Their tracking technique did not use cameras,
but is important to show the diversity of technology research The additional constraint needed for the estimation of the
for tracking and vehicle speed information. flow field is introduced in this method by assuming that the
flow (V x, V y) is constant in a small window of size m X m
Li et al. determined vehicle speeds by utilizing a CCD
with m > 1, which is centered at Pixel x, y and numbering the
camera and looking at the vehicle positions in video
pixels within as 1...n, n = m2, a set of equations can be found:
frames.[8] The speed is determined geometrically by the two
vehicle positions and their spatial relationship to the known
fixed CCD camera position. In our work, the spatial

www.ijcat.com 33
 International Journal of Computer Applications Technology and Research
Volume 2– Issue 1, 32-36, 2013, ISSN: 2319–8656

13. The video is then read and converted to frames.

. 14. Each image is converted from gray level to binary.
15. Reading each frame, the 2D coordinates of centroid for car
.
are extracted and stored in a matrix with corresponding
frame number.
where are the pixels inside the window, 16. The structure of the matrix is such that:
and are the partial derivatives of the Frame Number x - coordinate y - coordinate
image with respect to position x, y and time t, evaluated at
the point and at the current time.
17. The calibration parameters for the fixed camera
These equations can be written in matrix form , (calculated earlier) are used to convert 2D coordinates to
where 3D coordinates. That is, 2D Image coordinates are
converted to 3D World coordinates using calibration
parameters.[5]
A new matrix is obtained:

Frame X -coordinate Y- coordinate Z-coordinate

Number

The Lucas-Kanade method obtains a compromise solution by

The Euclidean distance between each successive matrix
the least squares principle. Namely, it solves the 2×2 system
element is calculated. Distance between 2 centroid P (xi,,yi,
or zi) and Q(xj,,yj, zj) in world space is calculated as
distij = √ (xi-xj)2 +(yi-yj)2+(zi-zj)2
Where n= total number of frames captured, i= 1 to n-1 and
where is the transpose of j=i+1 to n
matrix . That is, it computes 18. The total distance traveled (in millimeters) by the object
across the images is calculated. Total number of world
coordinates traveled (in mm) is given by:
n-1 , n
D = Σ distij
i=1, j=i+1
19. This distance is converted to actual distance traveled (in
with the sums running from i=1 to n. centimeters) using pixel to distance ratio (ctod) which is
calculated earlier (from preprocessing). The total distance
traveled by the vehicle under consideration is calculated
3.3 Steps Followed: as:
1. Input Video stream is captured. dtraveled = D * ctod
2. Convert AVI file from RGB to Intensity. Where dtraveled (in centimeters) is the total distance
3. The Intensity is then sent to Optical Flow block and traveled by the vehicle.
velocity vectors in the form of matrix are obtained. 20. The time of the motion (in seconds) is calculated using
4. The matrix is then sent to „Threshold and Region Filtering the following relation:
Block‟. ttraveled = n / fps
5. Inside the block there is a Velocity Threshold block which
Where n = number of frames for which motion of car was
calculates the mean threshold velocity and gives a binary
threshold image. studied,
6. This Threshold image is then divided into 2 halves using fps = frame rate of the AVI (number of frames per
Submatrix block and processed individually. The video second). This is obtained from the AVI information.
sequence is sent to Blob Analysis block. This block 21. Estimated velocity of the vehicle in centimeters per
calculates statistics for labeled regions in a binary image. second is calculated as:
These labeled regions are known as blob. Vel (cm/s) = dtraveled (cm) / ttraveled (s) [11]
7. The Blob Analysis block returns region of motion and the
coordinates of centroid of the moving objects in the video
sequence.
8. The Threshold Image (in Binary form), coordinates of
Bounding box and centroid are sent to „Display Results‟
block.
9. Centroid is superimposed on the original video using Draw
Markers block.
10. The video obtained in step (9) is subtracted from original
(divided) video to obtain image with centroid only.
11. The individual halves are stored in different AVI files.
12. The halves are concatenated to obtain the video sequence
with superimposed centroid.
www.ijcat.com 34
 International Journal of Computer Applications Technology and Research
Volume 2– Issue 1, 32-36, 2013, ISSN: 2319–8656

As a process of initialization, camera is calibrated and

coordinate to distance ratio is calculated. Further, the video
sequences were captured using a digital camera with 15 fps
sample rate and resizing the images to 120 X160 pixel
resolution.

The proposed method was first implemented for one

vehicle.[12] A camera was mounted on a height and the video
was captured. The calibration parameters were calculated
using Calibration toolbox. The pixel to distance ratio was
calculated and was stored for further analysis.

Fig 4.1: Original Video The experiment was repeated for multiple videos and 8
cars and the result is tabulated. Below are the output results in
the form of Speeds and motion vectors of cars.[9]

Table 4.1: Average Speed of Cars

cars 1 2 3 4

Speed(km/h) 1.22 1.25 1.28 1.31

cars 5 6 7 8

Speed(km/h) 1.27 1.30 1.35 1.32

Fig 4.2: Motion Vector

Table 4.2 : Motion Vectors of Cars

cars 1 2 3 4

Motion 0.3506 0.3548 0.3537 0.3586

vector

cars 5 6 7 8
Motion 0.3671 0.3672 0.3753 0.3798
vector

Fig 4.3: Tracking the vehicle 5. CONCLUSIONS & FUTURE WORK

The objective has been to detect moving objects
and thereafter, calculate the speed of moving Vehicles and
motion vector.[12] While earlier we worked with object-
intrinsic properties such as the centroid of a moving object
in order to make a probable prediction of its immediate
future motion, methods to detect a rectangular boundary
for the object, then used background subtraction Simulink
models and but didn‟t get fair output for multiple vehicles.
Further we made an attempt using the Lucas-Kanade
method. [13] Although that it does not yield a very high
density of flow vectors, Lucas-Kanade Algorithm is robust
in presence of Noise. Vehicle trajectory is shown in fig.5.
It is window based local method. Average angular error is
less in Lucas-Kanade algorithm. There exist fast and
accurate optical flow algorithms which can be applied in
Fig 4.4: Thresholded Video future. Hence, in future local and global methods can be
combined for requirement of dense flow estimate,
4. RESULTS preserve discontinuities and to make it robust to noise.
All the coding and matrix representations have been .[14]
implemented in MATLAB. The proposed method was
subjected to various experiments in order to check its
accuracy and feasibility.

www.ijcat.com 35
 International Journal of Computer Applications Technology and Research
Volume 2– Issue 1, 32-36, 2013, ISSN: 2319–8656

9. He, Z., Liu, Y., Yu, H., and Ye, X., E.F. ( 2008 ),
“Optimized algorithms for traffic information collecting in
an embedded system”, Proc. Congress on Image and
Signal Processing, Vol. 4, pp. 220-223.

10. S. Baker, I. Matthews, E.F.( March 2004 ), ”Lucas-

Kanade 20 Years On: A Unifying Framework”, IJCV,
Vol.56, No. 3, pp. 221-255.

11. Lazaros Grammatikopoulos, George Karras, Elli Petsa,

E.F.(November 2005), “Automatic Estimation of Vehicle
Speed from Uncalibrated Video Sequences”, International
Symposium on Modern Technologies, Education and
Professional Practice in Geodesy and related fields,
Sofia,pp.03 – 04.

Fig 5: Vehicle Trajectory using Lucas- 12. Savan Chhaniyara, Pished Bunnun, Lakmal D.
Seneviratne and Kaspar Althoefer, E.F.( MARCH 2008),
Kanade algorithm
“Optical Flow Algorithm for Velocity Estimation of
Ground Vehicles: A Feasibility Study”, International
Journal on smart sensing and intelligent systems, VOL. 1,
6 REFERENCES PP. 1.
1. Y. Wang, J.Ostermann, Y.O.Zhang, E.F. (1995), “Video
Processing and Communications”, Prentice Hall. 13. J.L. Barron, D.J. Fleet, S.S.Beauchemin, T.A. Burkitt,
E.F. (1992), “Performance of Optical Flow Techniques”,
2. Mehrubeoglu, M. and McLauchlan, L., E.F. ( 2006 ), Computer Society Conference on Computer Vision and
“Determination of traffic intensity from camera images Pattern Recognition, pp. 236-242.
using image processing and pattern recognition
techniques”, Proc. SPIE-IS&T Electronic Imaging, SPIE 14. A. M Tekalp, E.F(1995), “Digital Video Processing
Vol. 6063, 60630Q-1 -60630Q-12. Englewood Cliffs”, NJ: Prentice-Hall.

3. An, X., Qin, X. and Bao, H., E.F.( 2006 ), “Automatic and
robust classification of independent motions in video
sequences”, Proc. SPIE-IS&T Electronic Imaging, Vision
Geometry XIV, SPIE 6066, 60660B-1 – 60660B-8.

4. Yu, X.-D., Duan, L.-Y., and Tian, Q., E.F.( 2002 ),

“Highway traffic information extraction from skycam
MPEG Video”, Proc.IEEE 5th International Conf. on
Intelligent Transportation Sys., pp.37-42.

5. Anagnostopoulos, C.-N. E., Anagnostopoulos, I. E.,

Psoroulas, I. D., Loumos, V., and Kayafas, E., E.F.( 2008
), “License Plate Recognition From Still Images and
Video Sequences: A Survey”, IEEE Trans. Intelligent
Transportation Systems,9(3), pp.377-391

6. Garibotto, G., Castello, P., Del Ninno, E., Pedrazzi, P, and

Zan, G., E.F.( 2001 ), “Speed-vision: speed measurement
by license plate reading and tracking”, Proc. IEEE
Intelligent Transportation Sys. Conf. , pp.585-590.

7. Pelegrí, J., Alberola, J., and Llario, V., E.F.( 2002 ),

“Vehicle detection and car speed monitoring systems
using GMR magnetic sensors”, Proc. IEEE 2002 28th
Annual Conf. Industrial Electronics Society, pp.1693-
1695.

8. Li, Y., Yin, L., Jia, Y., and Wang, M., E.F.( 2008 ),
“Vehicle speed measurement based on video images”,
Proc. 3rd International Conf. Innovative Computing
Information and Control, pp.439-442.

www.ijcat.com 36