Jetson Nano faster for object recognition with GPU

Robotics & Edge Computing Jetson & Embedded Systems Jetson Nano
,
1

I have a Jetson Nano 2GB. I use it to recognize objects. I use the code from this page Simple object tracking with OpenCV - PyImageSearch

Which can be seen here: AIComputerVision/person_tracking.py at master · mailrocketsystems/AIComputerVision · GitHub

I run it with CUDA:
detector.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
detector.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)

Everything works and starts well. My problem is that the FPS is very slow despite the fact that it is with the GPU. It runs between 3 to 2 FPS … and sometimes it stays loading for a long time (with a very small resize so that it loads faster, if I set it large it is extremely slow).

If I use the Jetson Inference it looks big and runs a little faster. What’s the difference?

How can I improve my code?

2

Another question: can I make the jetson nano inference code more optimized? Example: that it does not show the video when I run it, or that the video that I receive is smaller or of an exact size.

4

Hi,

The GPU performance depends on the way of implementation.
Since Jetson has integrated memory, not all the frameworks have the optimal solutions for it.

It’s more recommended to use jetson-inference.
It uses TensorRT as the backend inference engine, which is more suitable for Jetson platforms.

Thanks.

5

I’m just going to try to use jetson-inference, but how do I implement TensorRT with opencv?

What happens is that my code is already working and if I use inference I have to add a lot of lines.

6

You need to check followings:

  • OpenCV supports CUDA
  • OpenCV is 4.2 or later.
    Since 4.2, DNN backend of CUDA supported.
8

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