Object Detection Using Frame Difference

Hi all,

I’m working on object detection using frame difference. Usually, when talking about object detection, CNNs, such as SSD, RCNN, and YOLO, are used and they return accuracies and positions (x,y,w,h). However I cannot run those models due to the hardware limitation (Jetson nano and in real-time) and also find them a bit unnecessary.
(I am aware of TensorRT optimization and its benchmark)

What I want to do is,

Frame difference -> cv2.findContours() or available contour algorithms -> detection by using sliding window + classification (ResNet50 in TensorRT) on each contour

I managed to write my code using cv2.cuda for frame difference part, but cv2.cuda doesn’t have findContours()

Is there any alternative for cv2.findContour?

Also, I want to hear some other approaches than sliding window technique.

I want to exploit every resource of Jetson Nano.

Thank you.


Unfortunately, we don’t have a CUDA API for image contour.
An alternative is to use the C++ version findContours() and inference it on TensorRT after memcpy.