Kalman filtering is better suited for tracking, not detection. Maybe the software you have doesn't provide you 'detection' and 'tracking' results separately. One of the recent versions of MatLab has a demo/example for exactly what you are looking for - it draws bounding boxes on people in a video and tracks the change of positions of the boxes over successive video frames; the code is well documented so you should easily be able to pull out the snippets relevant to 'tracking' and combine it with your 'detection' code.

Thanks for your reply. There is a question here... in situation that I did the detection part and have all the bounding boxes, how can I use those boxes for tracking. I mean, it is a bit weired to have all the boxes and then track them(what is the reason for tracking then?!). It should be like, we have only the initial boxes for the first image, then track them.

One solution can be simply using frame subtraction. Since you have the last positions of your boxes, you can relate them to the new positions obtained by the subtraction.

also by fusing the subtraction based estimation (or any other one) with the current positions in Kalman filter you would have better localization or tracking.

Finally, you can take a look to these papers:

robust pedestrian tracking using a model-based approach

Pedestrian Tracking by Associating Tracklets using Detection Residuals 

If you are performing detection on each frame, you could see this as an association problem, it means, to relate detected boxes within consecutive (or maybe non-consecutive) frames. In "Video Tracking: Theory and Practice" by Emilio Maggio and Andrea Cavallaro, (I don't remember the chapter) is presented a graph-based approach to solving this, in which the aim is to track different targets ("boxes" in your case) over a set of frames.

If you have boxes, use mean shift / cam shift algorithm. They are efficient and effective and there are plenty implementations of both.

I would also suggest to look into multi-clustered particle filtering. This software might be useful: http://www.dis.uniroma1.it/~previtali/?page_id=37, and it does exactly what you are looking for

http://www.dis.uniroma1.it/~previtali/?page_id=37

Hello Marjan,

A good idea for tracking multiple objects is to define this problem as a LSAP (Linear Sum Assignment Problem). Besides, the usage of Kalman filtering and appearance-based matching methods can help in cases where occlusions appear on scene.

You can find more information and an algorithmic explanation in this paper:

http://www.sciencedirect.com/science/article/pii/S0957417415004182

Some results can be visualized in this video:

https://www.youtube.com/watch?v=Rqk-vFKzAcQ

I hope that you can solve your problem soon :-)

Good luck!

Article Expert Video-Surveillance System for Real-Time Detection of ...

Hello Marjan Ramin,

Have you solved this problem as I am currently facing the same situation with you. As I have bounding boxes and know the start point of the frame. I would like to know how to track with these feature extracted from these bounding boxes. Could you please let me know how to solve?

It's been a long time since you asked but I will try to give an idea.

You can track the bounding boxes (bbox) based on the Intersection over union (IoU). You skip the first frame and start from the second one, calculating the IoU of each bbox over all bboxes so you will know that the higher IoU correspond to the same bbox.