Dept. of ECE 1 Feature-based Object Tracking Dr. Dapeng Oliver Wu Joint Work with Bing Han, William Roberts, and Jian Li Department of Electrical and Computer Engineering University of Florida
Jan 15, 2016
Dept. of ECE 1
Feature-based Object Tracking
Dr. Dapeng Oliver WuJoint Work with Bing Han, William
Roberts, and Jian LiDepartment of Electrical and Computer
EngineeringUniversity of Florida
Dept. of ECE 2
Outline
What is object tracking?
How to track an object?
KLT feature-based tracking algorithm
Our feature-based tracking algorithm
Experimental results
Conclusions
Dept. of ECE 3
What is Object Tracking?
Object tracking:
Track an object over a sequence of images
Dept. of ECE 4
Why Object/Vehicle Tracking is Challenging?
Occlusion or partial occlusion Re-appearance A vehicle makes a left/right turn A vehicle passes another vehicle in the same direction Tracking of a small imaged object (consisting of 4 pixels) Shadow effect: use local histogram-based equalization Cloud effect Tracking of many vehicles (e.g., thousands of vehicles in a city) Clutters Parallax problems (caused by high-rise buildings) Multi-camera based tracking (e.g., urban surveillance networks)
Dept. of ECE 5
How to Track an Object?
Kanade-Lucas-Tomasi (KLT) feature-based tracking algorithm
Han, Roberts, Wu, Li (HRWL) feature-based tracking algorithm
……
Dept. of ECE 6
KLT Feature-based Tracker
1. Identification of feature points
2. Determination of an optimization criterion for feature correspondence
3. Computational method to solve the optimization problem
Web: http://www.ces.clemson.edu/~stb/klt/References [1] Bruce D. Lucas and Takeo Kanade. An Iterative Image Registration Technique with an Application to Stereo Vision. International Joint Conference on Artificial Intelligence, pages 674-679, 1981. [2] Carlo Tomasi and Takeo Kanade. Detection and Tracking of Point Features. Carnegie Mellon University Technical Report CMU-CS-91-132, April 1991. [3] Jianbo Shi and Carlo Tomasi. Good Features to Track. IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. [4] Stan Birchfield. Derivation of Kanade-Lucas-Tomasi Tracking Equation. Unpublished.
Dept. of ECE 7
Identification of Feature Points
Harris criterion:
GGx 2tracedet kC
2
2where x x y
x y y
I I I
I I I
G
If threhold, declare as a feature point;
otherwise, declare as a non-feature-point.
C x x
x
For low threshold For high threshold
Dept. of ECE 8
Feature Correspondence
Determination of an optimization criterion for feature correspondence (find a similarity measure)
Correlation Information distance: e.g., mutual information
1
1 1 2 1 2
norm:
E.g., norm of [ , , , ] is (| | | | | |).TK K
l
l X x x x x x x
2
2 2 2 1/ 22 1 2
norm:
E.g., norm of is ( ) .K
l
l X x x x
Dept. of ECE 9
How to Solve the Optimization problem?
Key difficulties: Discrete optimization problem Gradient methods not directly applicable
Exhaustive search: incurs exponential complexity Lucas-Kanade computational method:
Newton-Raphson type search
Dept. of ECE 10
More about KLT Tracker
Hierarchical search Work for any dimensional vector Can address affine motion (translation, rotation, scaling) Can address intensity adjustment
Dept. of ECE 11
Hierarchical Search
Dept. of ECE 12
Limitation of KLT Tracker
Limitation: KLT tracker does not guarantee the corresponding point in the next frame is a feature point.
Why? Because KLT tracker only uses the Harris criterion for the first frame
but not for other frames.
What problem it may cause? KLT may not handle occlusion well.
How to improve it? Evaluate the quality of the corresponding point using the Harris
criterion Refer to:
B. Han, W. Roberts, D. Wu, J. Li, ``Robust Feature-based Object Tracking,'' SPIE Defense & Security Symposium 2007, Orlando, FL, USA, April 9–13, 2007.
Dept. of ECE 13
Our Tracking approach
Feature tracking: SSD criteria applied to find the feature point whose window minimizes the following energy function:
Feature quality: If the quality of a tracked feature point decreases below a chosen threshold, that point is removed from consideration. To compensate, new features are identified in the same window. The feature point with the minimum SSD criteria is retained if its SSD is below a certain threshold.
2 tyxIdttdyydxxIdydxEt ,,,,,
Dept. of ECE 14
Tracking Diagram
Feature point in current frame, new frame
Find point that min. SSD criteria
Meets threshold?
Retain feature
Identify new feature points
Select feature that min. SSD
Meets threshold?
Feature lost
Retain new feature
N
N
Y
Y
Dept. of ECE 15
Experimental Results
Biker image sequence
50 frame sequence of bikers.
Video contains global and local motions.
Objects undergo scaling, translation, and rotation.
Dept. of ECE 16
Experimental Results (2)
Biker image sequence (Cont’d)
1st frame. 258 features
selected.
50th frame. 241 features
tracked.
Dept. of ECE 17
Experimental Results (3)
Biker image sequence (Cont’d)
Blue: our approach. Red: KLT algorithm.
Dept. of ECE 18
Conclusions
Number of successfully tracked features increased by over 10% versus Tomasi-Kanade’s approach.
Computationally inexpensive and robust against various types of object motion.
For pure translational object motion, the combined algorithm does not offer improved performance to Tomasi-Kanade’s method.
Dept. of ECE 19
Future Research Direction
Design tracking schemes, which are proven to converge.KLT algorithm does not guarantee convergence; no
proof of convergenceIt is important to design convergence-guaranteed
tracking algorithms.We will systematically study this problem.
Dept. of ECE 20
THANK YOU!
THANK YOU!