Imagine a video from surveillance system and me as a developer of some algorithm for video summerization for this system.

There are two scenarios for me which I need help.

Imagine a long video that people commute in groups in front of camera and our task is to take the frames that "most of the faces are detected and of course detected faces are clear and recognizable".

The rigors over this task is that finding main frames (for example ->50) from 20000 frames which most of faces are detected and this demands evolutionary algorithms. I designed some PSO algorithm to sample data but I don't know how to make it smart enough to look for frames which faces are clear and detectable. (Do you have any idea, resource or code)

In the second scenario, imagine the people are coming one by one. so how can I keep each person passing the lobby in my summary (don't miss anyone, and also don't keep repetetive ones as much as possible).

Keep in mind that 2nd scenario will be added when I implement the face recognition system. So 1st scenario is more important.