I think that we can program our own classification model on live video or video file. The problem is if the speed of classification (on GPU) can match the frame rate.

Regards