Is there anyone able to track successfully?

[MuhammadAsadJaved]

Hi friends, Is there anyone able to track successfully? I have tried to contact the author on LinkedIn as well but he is not replying. seems like he is not managing this project anymore. I found one problem that the detector and tracker both working but when It go to the tracking part, the tracking is tracking but not drawing the bounding boxes on the right place.

These two pictures are from the same run. As you can see at the start the detector detects the object see the massage on window YOLO DETECTION and then passed to the tracker, the window also has massage GOTURN TRACKING and FPS information. So seems like the only problem is not drawing the bounding box after tracking. So we can try to figure it out if the author did not reply.

Thanks

[gzc-123456-gzc]

@MuhammadAsadJaved My friend, don’t I remember that you could not track it before? How to successfully track the drone now

[MuhammadAsadJaved]

@MuhammadAsadJaved My friend, don’t I remember that you could not track it before? How to successfully track the drone now

Hi Dear, These are the same old screenshots, The first image is from the start of the program when it detected a drone, after that the bounding box disappeared during the tracking(in the second image) but still there is text GOTURN TRACKING and FPS information. So I just noticed that it is tracking the drone but not drawing the bounding box. we need to figure out this problem.

• Here is the LinkedIn of the author. You can request him as well. Maybe hi resolve this issue for us. (https://www.linkedin.com/in/jonathan-boinet/)

[Boinet]

Hi guys thanks for your interest in the project!
From my experience when FPS goes above 30 and not bbox, it's a problem with the training of weights most probably yolo. I would suggest the following to debug the code:
1° shut the tracking off, and run only sliding detection window in non-real-time mode. This should make it clear whether it is a problem linked to the detection training.
2° Check the tracker itself to see what is the box input image and output box.

Good luck

[MuhammadAsadJaved]

Actually, Yolo is detecting very well, I have verified the detector with mainDetect.py

I will try again according to your suggestions. Thank you for your time and guidance.

[MuhammadAsadJaved]

Hi guys thanks for your interest in the project!
From my experience when FPS goes above 30 and not bbox, it's a problem with the training of weights most probably yolo. I would suggest the following to debug the code:
1° shut the tracking off, and run only sliding detection window in non-real-time mode. This should make it clear whether it is a problem linked to the detection training.
2° Check the tracker itself to see what is the box input image and output box.

Good luck

Hi Sir, @Boinet

I just figured it out the problem is with the tracker draw function.
https://github.com/tau-adl/Detection_Tracking_JetsonTX2/blob/master/DetectAndTrackUtils.py#L58

I just checked the x1 and x2 , y1 and y2 values was same, so tracker was not drawing bounding box. So i just added 50 and 30 in x2 and y2 to see the effect of the bounding box. Now it's drawing the bounding box but just after the detection ,the tracker's bounding box start moving at the bottom left. as you can see the first picture is detection and the second picture is very next frame after detection when it start tracking. So now I want to figure it out that why tracker's draw function is getting wrong values for bounding box. Which function is passing output to it? Do you have some problem to resolve this issue? As we already found the error so it may not take your too much time.

Update: Also checked for the tracker only

https://github.com/tau-adl/Detection_Tracking_JetsonTX2/blob/master/goturn/tracker/tracker_manager.py#L65

The bounding box values are same and far from object. So it might be weights problem. i will try it.

[MuhammadAsadJaved]

@gzc-123456-gzc Hi, It can detect but can't track. The tracker is writing wrong bounding boxes.

[gzc-123456-gzc]

@Boinet Hello friends, do you have a trained caffemodel file here

[gzc-123456-gzc]

@MuhammadAsadJaved Actually I think you can use Trim840.mp4 video to try its detection and tracking effect

[MuhammadAsadJaved]

Have you tried that video? @gzc-123456-gzc

[gzc-123456-gzc]

@Boinet Hello , do you have Trim840.mp4 video here

[gzc-123456-gzc]

@MuhammadAsadJaved Not yet, my environment configuration keeps going wrong. By the way, do you have Trim840.mp4 here?