Revision history [back]

Some ideas:

Take only a frame from the video, like a photo. It work for me.

Try houghcircles approximations instead of rectangle for contours, this is an example https://docs.opencv.org/master/d1/de6/samples_2cpp_2tutorial_code_2ImgTrans_2houghcircles_8cpp-example.html

Try using lights with green filters, almost all cameras work better with green lights.

I'm also working with object recognition for logistics procedures. Be in touch.