Tracked boxes have delay related to real objects coordinate in NvDCF tracker?

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

• Hardware Platform (Jetson / GPU):jetson
• DeepStream Version:5.1
• JetPack Version (valid for Jetson only):4.5
• TensorRT Version:7.1.3
• Docker image:5.1-base

Hi,
I use NvDCF as tracker in pipeline that have one pgie and one sgie, and I linked the tracker after pgie for tracking the objects. and I also use deepstream-5.1 base docker image.
My problem is that the all of detected boxes have delay related to real objects coordinate when I use dcf tracker, but with IOU tracker every thing is good.
Note that When the speed of objects are high this situation is occured but when the objects passing slow in scenes even with dcf tacker is works correctly. The total FPS pipeline is about 150 FPS.

pipeline is like this:

sources >> decodebin >>streammux >> pgie >> queue >> tracker >> queue >> sgie >> streamdemux >> … >> fakesink

tracker.yml (6.6 KB)

3

Can you elaborate on this? How did you observe the delay? The object id is wrong?

4

@Fiona.Chen,
No the object id is correct.
The results are like this:
Almost the most of coordinated boxes are incorrected, but with IOU tracker are good.
As a mention above, this situation is occurred with DCF tracker when the objects are fast relatively.

1
11440×960 138 KB

5

The bbox is predicted by DCF but not detected by PGIE. The picture you posted shows the prediction does not match to the actual bbox.

Can you share your NvDCF ll config file?

6

@Fiona.Chen ,
The example shows two scenarios:
Box 1 : When the objects passing from top to down the box 1 is occurred.
Box 2 : When the objects passing from down to top the box 2 is occurred.
These results show the predicted boxes have delay relative to real boxes.

1 (1)
1 (1)1440×960 139 KB

tracker.yml (6.6 KB)