8. Hardware in Real-time Tracking
• MEMORY
Important Tracking system encountering limited memory
problems.
• FRAME RATE
~30 FPS
• PROCESSORS - DSP
• Allow saturated arithmetic operation
• Powerful operation ability
• Can do several memory accesses in a single instruction
9.
10. Object Detection and Tracking
• In a video sequence an object is said to be in
motion, if it is changing its location with
respect to its background
• The motion tracking is actually the process of
keeping tracks of that moving object in video
sequence i.e. position of moving object at
certain time etc.
11. Flow Chart
Idle
Image
acquisition
Object
Detection
Image
acquisition
Object
tracking
Object
No
Lost?
Y es
12. Method 1: Absolute Differences
= Image subtraction
D(t)=I(ti) – I(tj)
Gives an image frame with changed and
unchanged regions
Ideal Case for no motion: I(ti) = I(tj), D(t)=0
14. Results:
Frame1 Frame10
Difference of Two Frames
15. Absolute Difference
Methods for Motion Detection
Frame Differencing
Background Subtraction
Draw Backs:
involves a lot of computations
Not feasible for DSP implementation
17. Image
128 26 125 243 87 Signature Vectors
96 76 43 236 125
10110101
00101011
128 129 235 229 209 Signatur vector generation for
.
all pixels
228 251 229 221 234
.
.
227 221 35 58 98 10111010
List 10110101
Generation List
00101011 population
.
.
.
Generated List 10111010
Signature vector
matching
18. Census Transform:
Advantages:
Compare only two values 0 or 1.
Similar Illumination Variation for pixel and
neighbouring pixels
Draw Backs:
As we only deal with only 0`s and 1`s, this method is
sensitive to noise.
Calculate, store and match process computationally
Expensive
20. Morphology Based Object Tracking
• Image Differencing
Background • Thresholding
Estimation
• Contours are registered
Object • Width, height and histogram are recorded for each contour
Registration
• Each object represented by a feature vector (the length,
Feature width, area and histogram of the object)
Vector
22. Morphology Based Techniques
Advantages:
Can Track Multiple objects Objects are registered
based on their anatomy
Helpful for Object Merging
Draw Backs:
Object registration complex and slow process
For multiple object registration per frame more
complex
23. Method 4: Lucas-Kanade Technique
• Visual motion pattern of objects and surface in a
scene by Optical Flow
Frame 1 Frame 2
24. Method 5: Mean shift
• An algorithm that iteratively shifts a data point to the
average of data points in its neighborhood
Choose a search window Compute the MEAN
size in the initial location location in the
search window
Repeat until Center the search
convergence window
at the mean
25. Intuitive Description
Region of
interest
Center of
mass
Mean Shift
vector
Objective : Find the densest region
Distribution of identical balls
26. Intuitive Description Region of
interest
Center of
mass
Mean Shift
vector
Objective : Find the densest region
Distribution of identical balls
27. Intuitive Description
Region of
interest
Center of
mass
Mean Shift
vector
Objective : Find the densest region
Distribution of identical balls
28. Intuitive Description
Region of
interest
Center of
mass
Mean Shift
vector
Objective : Find the densest region
Distribution of identical balls
29. Intuitive Description
Region of
interest
Center of
mass
Mean Shift
vector
Objective : Find the densest region
Distribution of identical balls
30. Intuitive Description
Region of
interest
Center of
mass
Mean Shift
vector
Objective : Find the densest region
Distribution of identical balls
31. Intuitive Description
Region of
interest
Center of
mass
Objective : Find the densest region
Distribution of identical balls
33. CAMSHIFT
--Continously Adaptive Meanshift
Modified to adapt dynamically to the colour
probability distributions
More real time
For each frame-> MEAN-SHIFT is
applied with several iteration
Store the location of the mean and
calculate new window size for next
frame
34. New development
• Combine with different features. SIFT features,
colour feature & texture information
• Camshift algorithm combined with the Kalman
filter.
35. Result
Arithmetic and Time taken
Algorithm Logic by
operations Algorithm
Absolute
4230100 16
Differencing
Census
2416000 5. 4
Transform
Morphological
352210 14.2
Tracking
Kanade Lucas 500825 0.486
36. Comparison
Computationally
Easy to implement
expensive
Absolute Differences Allows continuous Slow and low
tracking accuracy
Computationally
expensive
Census Transform Immune to noise and
Illumination changes Complex if
Multiple objects
per frame
Can track multiple Slow
Feature Based
objects well
Large Memory
consumption
37. Comparison
High accuracy
KLT
Less execution time Large memory
Robust to noise and
dynamic scene
Ineffective if
Computationally less there is heavy
MeanShift & CAMShift expensive occlusion
38. Conclusion
• KLT algorithm has the best performance with
higher accuracy and less computation time
• It requires combination of methods to achieve
the appropriate object detection and tracking
according to the proposed scenario
39. References
• S. Shah, T. Khattak, M. Farooq, Y. Khawaja, A. Bais, A. Anees, and M. Khan, “Real Time Object
Tracking in a Video Sequence Using a Fixed Point DSP,” Advances in Visual Computing, pp. 879–
888.
• K. Huang, L. Wang, T. Tan, and S. Maybank, “A real-time object detecting and tracking system
for outdoor night surveillance,” Pattern Recognition, vol. 41, no. 1, pp. 432–444, 2008.
• J. Li, F. Li, and M. Zhang, “A Real-time Detecting and Tracking Method for Moving Objects Based
on Color Video,” in 2009 Sixth International Conference on Computer Graphics, Imaging and
Visualization. IEEE, 2009, pp. 317–322.
• W. Junqiu and Y. Yagi, “Integrating color and shapetexture features for adaptive real-time
object tracking,” IEEE Trans on Image Processing, vol. 17, no. 2, pp. 235–240, 2008.
• Q. Wang and Z. Gao, “Study on a Real-Time Image Object Tracking System,” in Computer
Science and Computational Technology, 2008. ISCSCT’08. International Symposium on, vol. 2,
2008.
• Y. Meng, “Agent-based reconfigurable architecture for real-time object tracking,” Journal of
Real-Time Image Processing, vol. 4, no. 4, pp. 339–351, 2009.
• [Y. Yao, C. Chen, A. Koschan, and M. Abidi, “Adaptive online camera coordination for multi-
camera multi-target surveillance,” Computer Vision and Image Understanding, 2010.
Image/appearance based trackingThere are wide application in real time object detection and tracking.
Multiobject (people) tracking within a video of a pedestrian passageway. The dynamic motion vectors attached to each individual represents direction of movement and speed.
Here, we can see how a mobile robot can detect and track this red ball. It moves accordingly to the red ball movement.
As the objects move over time, ther are different illumination and motions of small objects; due to perspective, occlusion, interaction between objects and appearance or disappearance of objects.2. to track targets and keep on detectingnew ones on a moving camera platform at the same time,the traditional motion detector based on the backgroundsubtraction can not be applied here.3. How to...In the same time..4. Because of comprehensive search ,it is hard to meet strict time constraint in real-time
To solve these problems...The choice of hardware can increase the performance of object detection and traking in real time which has hard time constraintsMemory is....There are several processor, such as FPGA depends on the choice of soft processorASICulfill the speed criteria of real-time, but it is complicatedGPU increase the speed up the computation at the bottom level method (optical flow)
The first term is proportional to the density estimate at x computed with the kernel G. the second term is the mean shift. This part is the mean of the window. We calculate it by using a kernel function, which gives different weights to all points inside the window. the mean shift vector thus always points toward the direction of maximum increase in the density.
http://www.codeproject.com/KB/GDI-plus/MeanshiftTracking.aspxChoose a search windowsize.2. Choose the initial locationof the search window.3. Compute the mean location(centroid of the data) in thesearch window.4. Center the search windowat the mean locationcomputed in Step 3.5. Repeat Steps 3 and 4 untilconvergence
Camshift is based on mean shift, but the window size is changed in in video sequence, when the object moves, the size and locations of color distributions will change over time. Mean shift. mean shift algorithm uses fixed window size. So it might fail. However, camshift can deal with this problem by adjusting the window size according to the distribution. Real time
3. Firstly, initialize the Gaussian mixture model to get the background image, and then using the background differential with the current frame to detect the moving objects. Kalman filter is used to predict the centre of the searching window in the next frame. Then, camshift will find the optimum position of the target in order to modify the prediction. This can improve the speed of camshift algorithm and solve the occlusion problem.
Theses results are from the paper by Shah, in which he has demonstrated the results of the all the above mentioned methods for the same task of tracking a moving ball with DSP hardware.As we can see that the Lukas kanade is the fastest compared with the traditional methods.
The absolute difference and census transform are easy to implement but computationally expensive and slow. Feature based method can track multiple objects, but it is also slow.
KLT algorithm is can detect objects fast and accurately and it is robust to noise and dynamic scene. but it requires large memory, when the search window size is large.Mean shift has low computation cost. But it might fail in case of heavy occlusion and it can only detect single object. This can be solved by combining different algorithm, for example, SIFT feature descriptor and Kalman filter.