MobileNet SSD and tracking about multitarget-tracker

[Nuzhny007]

Hi!
Thank you for your attention to the project.
In latest version to starting MobileNet example you need to write:
./MultitargetTracker your_video.mp4 -e=4

P.S. An user manual or paper coming soon...

[KuntimaKiala]

Hi
Thanks for your time, as I mentioned, I did run the project, since I was interested with the MobileNet, I used ./MultitargetTracker your_video.mp4 -e=4, and it worked fine, but I just did not understand how you connect the detector with the tracker, it seems like you run the detector forever, and you accord one tracker to each person that appears in the scenario, but the detector does not know the differences between persons, so can you tell me, how do you track each individual ? without putting many tracker in the same person, what is your condition for tracking and stop tracking ? you have the tracker and the detector running in parallel ?

[Nuzhny007]

Yes, Detector - MobileNet (or YOLO, or HOG, or Haar, or background subtraction etc) works on each frame. We get a list of the raw detections - regions (or blobs, or bounding boxes). Here: https://github.com/Smorodov/Multitarget-tracker/blob/master/VideoExample.h#L310

For raw detections on each frame we solve a spatio-temporal data association problem. Global data association - Hungrian algorithm (or Kuhn–Munkres algorithm, or Munkres assignment algorithm). Also it can use max flow in weighted bipartite graphs. Here: https://github.com/Smorodov/Multitarget-tracker/blob/master/VideoExample.h#L330
Local tracking and trajectories processing - Kalman filter.

But modern algorithms use another approaches for global data association:

1. http://vision.cse.psu.edu/courses/Tracking/vlpr12/lzhang_cvpr08global.pdf
2. http://www.milanton.de/dctracking/
3. https://github.com/nwojke/mcf
   I want to include the latest library to the our project.

[KuntimaKiala]

Thanks,
So if I understand You, You use the Hungrian algorithm to assign each person to one tracker according to they spatial-temporal position ? I guess I have to check first what is spatio-temporal data association problem, can you show me where do you use the Hungrian algorithm in your project ?

[Nuzhny007]

You use the Hungrian algorithm to assign each person to one tracker

...to one Kalman filter.
Or to one tracker like KCF if this object wasn't found on this frame.

Hungrian use here: https://github.com/Smorodov/Multitarget-tracker/blob/master/Tracker/Ctracker.cpp#L73

[KuntimaKiala]

Thanks for your time, I'll look into it...