liveposetracker icon indicating copy to clipboard operation
liveposetracker copied to clipboard

Published 20 hours ago • verified publisher icon ortegatron

Occluded pedestrian detection

Open xiaobaozi1996 opened this issue 5 years ago • 4 comments

Sorry to bother you again. Recently, I tested the code for my own dataset, but in the case of occlusion, ID is either changed or transferred. What parameters should I modify to avoid.

xiaobaozi1996 avatar Aug 08 '19 14:08 xiaobaozi1996

For sure it's not a perfect algorithm. You can try to twerk parameters on kalman_filter.py

ortegatron avatar Aug 08 '19 16:08 ortegatron

Ok, I'll try it. By the way, Is there an effective method for tracking in the case of occlusion based on you knowledge

xiaobaozi1996 avatar Aug 09 '19 01:08 xiaobaozi1996

Hi @ortegatron - I'm having a similar issue where a track (ID) can be "stolen" if a person crosses in front of another, even just a little bit. Do you have any suggestions for the parameters to adjust that could help this happen less? I don't understand all of the math and such in the kalman_filter.py.

tilink avatar Sep 21 '19 17:09 tilink

I guess you need to ponderate more the appearance matching.

ortegatron avatar Oct 10 '19 03:10 ortegatron