Determining frequency of repetitive mouvements in live video feed

I want to analyse live video feed of people doing repetitive motion.
Specifically the will be a phone in a tripod looking at people doing indoor paddling. I want to determine the stroke rate.

The frame below is an example of what the camera will be looking at:

Sometimes the background can be a bit more messy and the view may not be as straight on such as this:

.

The typical frequency of the stroke would be anywhere between 40 and 80 stokes per minute. That is the rate i am trying to detect live.

Does anybody have a suggestion of what king of processing of the video stream i can perform to try to detect this frequency?

fairly easy if the camera is static relative to the people.

one idea: apply pose estimation, track their poses over time. you get a skeleton per human. then you’d pick a suitable joint, extract its coordinates over time and continue from there.

there are surely other approaches. this one is probably easy to get done.

1 Like

Thanks. Got the pose estimation going and i am extracting the wrist positions over time. Applying some Kalman filtering to smooth things out, detecting peeks and coming up with a frequency. Not perfect yet but getting there.