0

I have been working on a truck detection model using YOLOv4 darknet. The model performance is good. I have applied deep_sort object tracking to track the activity of the trucks in that region.

The problem with this approach is that the truck identity changes when it steers around showing a different side against the camera feed or is obstructed by another object.

Is there a way to make sure that truck ID does not change?

Link to a demo inference video

I have trained the model specifically for this video. Object detection works fine but tracking id changes.

Abhishek Purandare
  • 53
  • 1
  • 5
  • Look into re-identification methods. – B200011011 Feb 09 '22 at 11:27
  • Correct me if I'm wrong but the re-identification likely won't be useful as vectors for the back and the front side of a truck will be different. Thus, the similarity score will be lower. @B200011011. If you can share any links on this approach. It would be great. Thanks for the response! – Abhishek Purandare Feb 10 '22 at 08:33
  • I could not see your video. You can find many resources with re-id on google. You may also want to try different trackers that does allows better tracking under occlusion and perspective change. Someone else may be able to provide better answer. – B200011011 Feb 10 '22 at 10:13

0 Answers0