I am working on this project as my TIPE for the french "Concours aux grandes écoles d'ingénieur".
The aim of the project is to suggest an algorithm able to detect violent behaviors.

The video input must have a fixed background with the least as possible moving objects (except pedestrians).
The detection is based on the YOLOv5 network.
In order to track people on the video I am using a kalman filter.
The algorithm is not able to detect dangereous behaviors yet.

The red line is an estimation of the orientation of the arm, the pink dot is an estimation of the position of the hand.
The green got is the position of the head.

I have decided to use a YOLOv5 network to detect people because the previous detection method I was using was not persistent enough. For instance when changing video input the detection system was not able to adapt to the new dimensions of the video. Maybe I could find a way to adapt my old system unfortunately I do not have enough time to work on that.