2

I have only little experience with embedded systems. I have a Raspberry Pi 4 with 1gb RAM and a Camera Module v2.1 and want to do object detection and tracking on it. By best solution I mean a solution that is as accurate as possible and has as many FPS as possible.

I did some research on the internet, there are some suggestions, like

  • OpenCV DNN module, and then e.g. Yolo, ResNet
  • Tensorflow Lite
  • Darknet for Yolo, with NNPACK

Are there any papers which compare different solutions for object detection, which show how accurate and fast the algorithmns perform on embedded devices? Or do you have some experiences? Thanks!

1 Answer 1

1

OpenCV's DNN module is very optimized but it doesn't support quantized models at the time of writing this answer. Go with Tensorflow Lite on 64bit OS. I could get 15 FPS on Raspberry Pi 4. By the way, I have provided a C++ toolbox for object detection on ARM CPUs. Currently, only cascade classifiers are supported.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.