Tracker/Inferences Losing Object for One Frame Leading to Line Crossing Counting Issues

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hardware Platform (Jetson / GPU):

  • Target system: Jetson AGX Xavier
  • Currently testing on an x86 server

• DeepStream Version:
6.3
• Issue Type:
Bug / Tracking Inaccuracy

  • Module / Plugin:
  • NvTracker/NvDCF Visual Tracker
  • Function Description:
    We are trying to count people using our renset34 model. However, we are facing an issue where a single missed detection frame leads to incorrect counts. In all other frames, we achieve high detection confidence and valid tracking confidence. We have attached an image showing an excerpt of three frames from our counting process. It clearly demonstrates that an object is lost in one frame and then immediately re-detected in the subsequent frame, receiving the same tracking ID upon re-detection.

Any insights or recommendations to help mitigate this issue would be greatly appreciated.

We would appreciate any insights or recommendations regarding this behavior. Has anyone experienced similar issues with transient object loss? Are there any configuration adjustments or workarounds that can help maintain the object track even if it’s briefly undetected?

Thank you in advance for your assistance and suggestions!

Best regards,

3

I can see the object in frame 2 is clear and the detection confidence is high. Why the object isn’t detected in frame 2? Can you fix this detection issue firstly?