Accuracy and Performance issue for Object detection and tracking

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

• Hardware Platform : Jetson Nano
• DeepStream Version 5.0

Hello there!
I have created an application on top of deepstream-test2-python-app for vehicle counting. Ofcourse I have to use a model for object detection and a tracker to track the objects.
I am facing the following issues.
1 - While using the DetectNetV2_ResNet10, it is giving the good fps with smooth output. But the detection results are not very good especially if I test it on the videos other than provided in the sample videos.
2 - While using the TrafficCamNet_ResNet18 with KLT tracker, it is giving relatively better detection results for vehicles. But:
- To make the output video smooth, I have to skip the frames(i-e set interval > 0)
- The tracker accuracy is very low.
3 - While using the TrafficCamNet_ResNet18 with NvDCF tracker:
- The accuracy of tracker is good, but its computatinally very expansive and gives very low fps.
So, I need suggestions to make the application robust and accurate. What should be the way forward to achieve better accuracy/results/performance on Jetson Nano?
Thanks.

3

Hi,

Is it possible to retrain the model?
Since the use case is vehicle counting, you may get a better accuracy if fine-tuning the model for vehicle.

Below is our toolkit to help user customize the model:
https://developer.nvidia.com/transfer-learning-toolkit

Thanks.