search

Smorodov / Multitarget-tracker

Multiple Object Tracker, Based on Hungarian algorithm + Kalman filter.
Apache License 2.0
2.17k stars 647 forks source link

Integration of GMPHD filter #125

Open Nat-zewge opened 5 years ago

Nat-zewge commented 5 years ago

@Nuzhny007 you asked a question https://github.com/blefaudeux/gmphd/issues/2 about integrating GMPHD into your framework,i think its great for comparison, did you succeed? thanks for you great work by the way

Nuzhny007 commented 5 years ago

The right link: https://github.com/blefaudeux/gmphd/issues/2

And yes - I tried to use GMPHD, the old branch: https://github.com/Nuzhny007/Multitarget-tracker/tree/gmphd Results with snow: https://www.youtube.com/watch?v=DfZsMIBAmCc It can be seen badly, but the objects are held much worse than with our standard tracking. Therefore, I abandoned this path. If you have some ideas then I can implement they.

Nat-zewge commented 5 years ago

Yes got it. in theory, several papers say that a family of filters called Random Finite Set(RFS) filters(which GMPHD is one of them) perform better than other methods. perhaps you can have a look at this one: https://github.com/nucleusbiao/Pedestrian-Tracking-using-SMC-LMB-with-OpenCL
it's called labelled multi Bernoulli filter(it is another of the RFS filters), seems to work ok on the demos, but i haven't tried it on other data sets. have a look maybe to see how it works. it is worth integrating them into your framework, i am quite eager to test and compare with other methods.

Nuzhny007 commented 5 years ago

Thank you for link! It will be not very simple and fast but I'll try to implement this approach.

Nat-zewge commented 5 years ago

cheers 👍

Nuzhny007 commented 5 years ago

Hi! I'm tested LMB tracker but it too slow for real time work. I think it makes no sense to include it in this repository for the following reasons:

  1. It is slow.
  2. After testing on several videos I did not see a gain in accuracy.
  3. It need to add new dependency - armadillo library. Now projects depends only from OpenCV.
Nuzhny007 commented 5 years ago

What do you think about this approach? I'm reading this dissertation from nwojke : https://kola.opus.hbz-nrw.de/frontdoor/index/index/year/2018/docId/1708 Interesting! But it not seems that it can use for real life applications. Approach from "FollowMe: Efficient Online Min-Cost Flow Tracking with Bounded Memory and Computation" looks more realistic.

Nat-zewge commented 5 years ago

Ahh, yes it is slow because it uses particle filters for implementation, well if it does not give any marked improvement then there is no point in adding it, you are right. about the dissertation, why do you think that it won't work with real world applications? do you think it will be slow as well? the author seems to have shown results that it works in real scenarios.
https://arxiv.org/pdf/1611.06011.pdf this version also claims to do tracking online, the mathematics needs time to understand well, but the results seem ok, based on the table of comparisons(Table 1). i will check the second paper you sent, Thanks!!