Detection + classification + Tracking code with Deepstream sdk

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

• Hardware Platform : Jetson Nano 4gb developer kit
• DeepStream Version : 5.1
• JetPack Version : 4.5.1
• TensorRT Version : 7.1.3.1
• CUDA Version : 10.2.89
• Python Version : 3.6.9 , 3.10

I have ONNX model trained on YOLOv6. I have complete code that do detection and classification using ONNX model and do tracking using deepsort but while running it normally on jetson nano it is giving me 1.5 fps. I want to increase it to 25 fps, so i want to include deepstream framework in my existing code so that i can use jetson nano GPU as well.
Please guide me some way to proceed .

Thank you in advance.

3

Please test the models’ performance with trtexec tool to identify whether the GPU accleration can improve the whole performance.

The DeepStream SDK user manual is Welcome to the DeepStream Documentation — DeepStream 6.2 Release documentation

You may try to use deepstream-app sample to integrate your detector, classifier and tracker. DeepStream Reference Application - deepstream-app — DeepStream 6.2 Release documentation

4

May you tell in more detail version how i can integrate my custom model for detection and tracking. I reviewed that link but didn’t understand much.

5

If you want to integrate your models with Gst-nvinfer — DeepStream 6.2 Release documentation, please read the document first. It is hard to skip the details in the document to make you understand how the elements and pipeline work.

You can read deepstream-app document and try the samples to understand how the app and configurations work.

6

Hello @Fiona.Chen,
I have the understanding of how deepstream work. I have worked on deepstream-rtsp-in-rtsp-out, deepstream-test1 and other samples as well. But while building project with this deepstream pipelines I always used the model that were provided by deepstream for detection and tracking.
But now i want to use my custom YOLOv6 ONNX model for detection and the deepsort model for tracking using deepstream.

Please help me with some way to integrate this custom model
with deepstream. So that i can improve my fps.

7

The first suggestion is to upgrade your DeepStream version to latest 6.2 version. The latest gst-nvtracker has supported DeepSort tracking. Gst-nvtracker — DeepStream 6.2 Release documentation

The second suggestion is to read the sample of NVIDIA-AI-IOT/yolo_deepstream: yolo model qat and deploy with deepstream&tensorrt (github.com), there are YoloV4 and YoloV7 samples here. It may be a reference for the Yolo models.

8

Thank you for your help. I will try and let you know

9

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

10

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