- initialise an empty array to store all the data
- For each video:
- parse the labels.json file
- parse the fps/yolov7.json file
- for every data_augmentation:
- for each frame:
- initialise a dictionnary for every unique skeleton as they will require separate rolling windows (only if multiple people on the frame)
- if the offender skeleton is tracked
- add it to the rolling window (or create the rolling window is the skeleton id didn't exist)
- process it (velocity, velocity centre, angles)
- store it
- save the file