Feature/body tracking

[nathantsoi]

2 changes in this branch / pr:

• enable body tracking by default, so no extra build-time params are required
• add the body track id, so each body can be uniquely identified

[lpasselin]

Hi and thanks for the PR!

I always assumed body_id was simply an index. After seeing your PR, I assume body_id is unique and matches a specific person only.

It is sad the Capsule.body_index_map uses indexes instead of body_ids. Otherwise, we could have used a dict to store the body_skeleton without having Capsule.body_ids

Do you think we should save the capsule usage by replacing the body_id numpy array with a tuple? This tuple would be created without having to malloc and free in the c extension (see here: https://stackoverflow.com/a/55987196).

It's just an idea, let me know if you think it would be faster this way.

[romilpunetha]

Any ideas when this will make it to master?

[lpasselin]

It's not my top priority right now. If you absolutely need this branch and the newer features in dev/master, I'll try to find some time to work on it.

[romilpunetha]

It's not my top priority right now. If you absolutely need this branch and the newer features in dev/master, I'll try to find some time to work on it.

Yup. I needed the body tracking sdk for python. I'm trying to build something on top of it which will run on linux, hence needed this capability.

[0xc1c4da]

@lpasselin please find some time to work on it, I also need this capability, currently using @nathantsoi 's branch

[ZhengyangSun1020]

@lpasselin please find some time to work on it, I also need this capability, currently using @nathantsoi 's branch

Yep, I also need this haha.

[ZdenekM]

It would be great to have body tracking in the master!

[Jakub-Dv]

In my testing, I've found that cleanup functions capsule_cleanup_body_ids and capsule_cleanup_body_data are never called. What I found:

• no circular reference on body_skeleton
• len(gc.get_referrers(capture._body_skeleton)) == 1, as it should be
• when adding fprints to other cleanup functions, I can see that capsule_cleanup_capture is properly called every time capture goes out of scope

How to reproduce: Simply run python3 examples/pose_display_2d.py and watch process memory go up.

I don't really know why is that happening but it causes huge memory leaks (~14 MB/s if capture.body_skeleton is called on each capture and FPS is 30). @lpasselin Can you find some time to investigate this issue please?

Edit: I've managed to get body_ids and body_skeleton working (I've removed body_index_map as it was not working with the same 'fix'). See https://github.com/Jakub-Dv/pyk4a/tree/feature/body-tracking. I don't really know if this is the correct fix, but it works.