I am tracking multiple pedestrians walking in the street. However, there is a problem when the algorithm looses the track causing broken tracks. I am looking for a method to have a smooth track of each pedestrians. Does anyone has a suggestion?
Hello
You have asked a question that appears simple but is difficult to answer comprehensively. This is a fundamental problem in most object tracking systems. You can try different tracking algorithms (have a look at the attached paper for a start), but if that doesn't solve the problem you have to undertake a secondary processing stage to join broken tracks. This can be set up as a hypothesis testing problem and there are a few papers on it but not that much really. The basic idea is to test the end of one track with the start of another to see if they correlate in time and space. The problem is more difficult if there are multiple objects or occlusion.
Article Taxonomy of multiple target tracking methods
Hello,
For broken tracks you have to consider the end and start of broken part. Then need to check the correlation the start and end part considering the spatio-temporal factors. You can use KSP or POM algorithm. These will give frame independent detector working together with a joint optimization method. It is unaffected by occlusion and provide you real time analysis.
http://www.cs.berkeley.edu/~malik/papers/coiffmanBMM.pdf
Thank you for the paper, Mr Chowdhury. This is clearly a relevant paper for what I call extended object tracking (cars in a video sequence). There is however an underlying generic problem for point objects (occupying only a small number of pixels). This problem has not been looked at extensively AFAIK although there are a few references. I have seen some work on this problem in both radar and in sonar for multiple object contexts where the appearance of a new track can conflict with a break in an existing track. This is the fundamental problem to solve.
Thank you so much for your helpful answers. I considered the end of each track and found the closest bounding box in the next frame which is start point of a new track. I used euclidean distance to find the closest center of the next bounding box to the center point of the current bounding box. I defined a threshold to consider just those bounding boxes which are in the defined threshold area. It has solved the the problem in some parts. But I am not sure that it is precise and strong enough for the whole problem.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing