Optimizing SiamRPN++ and Siamese Tracker for Speed on Jetson Orin NX

Accelerated Computing Intelligent Video Analytics DeepStream SDK
, ,
1

• Hardware Platform: Jetson Orin NX

• DeepStream Version: 6.3

• JetPack Version (valid for Jetson only): 5.1.3

  • TensorRT version: 8.5.2.2
  • CUDA version: 11.4

I am working with the SiamRPN++ model for object tracking and am looking to use it on a Jetson Orin NX. Specifically, I am interested in integrating the Siamese Tracker, a SOT (Single Object Tracker), into DeepStream for real-time performance. I plan to use the pysot library, which can be found on this GitHub repository, to implement the tracker.

I am considering two approaches:

  1. Adding the Siamese Tracker as a custom plugin to DeepStream using Python.
  2. Optimizing the speed of SiamRPN++ within the DeepStream pipeline or via other methods.

I want to know if it’s possible to integrate the Siamese Tracker (SOT) from pysot as a custom plugin in DeepStream for efficient object tracking. Specifically, I am interested in how to do this with Python and how to improve the tracking speed for real-time applications.

Is there a way to:

  • Implement the SiamRPN++ model from pysot efficiently on Jetson Orin NX using DeepStream?
  • Integrate the Siamese tracker (SOT) as a custom plugin to leverage GPU acceleration?
  • Optimize the speed of this tracker to run smoothly in real-time, especially at high frame rates?

Any tips on optimizing the DeepStream pipeline or other ways to speed up the tracker are highly appreciated.

Thank you!

3

We have this guide to implement a custom tracker:Gst-nvtracker — DeepStream documentation
But it is C++ API to implement a custom tracker in DeepStream.
Can you share why current nvtracker isn’t enough for your project?

4

The existing nvtracker module doesn’t fully meet my project’s requirements for these reasons:

  1. Frame Processing Rate: I need high-speed processing focused on a single object, but nvtracker options are designed for multi-object tracking, adding unnecessary operations that slow down performance.
  2. Single Object Tracking: My project requires dedicated, single-object tracking, which the current nvtracker algorithms don’t support effectively. While the IoU tracker has the necessary speed, it lacks the precision required.

A Siamese-based tracker, however, would address both speed and accuracy for single-object tracking, making it a more suitable choice for my needs.

5

Regarding a single object, it depend on the detector. If the detector only output one object, nvtracker will only tracking the detected one object.
Regarding performance and accuracy, please check this doc: Gst-nvtracker — DeepStream documentation

6

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

7

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.