Real-time exercise form analysis with MediaPipe , looking for advice

For one month, I would keep this much smaller than it sounds. MediaPipe Pose and joint angles are a good choice, but I would not start by training a classifier for many exercises and “correct vs incorrect” form. That sounds simple on paper, but in practice the hard part is the video pipeline. Camera angle, bad lighting, people going partly out of frame, noisy landmarks, different body shapes, blah blah.. I would pick one exercise first, maybe squat. Get the phone camera working live, draw the pose, compute the main angles, and make sure the numbers are stable while someone is actually moving. That alone will teach you most of the problems. For form feedback, simple rules are probably enough for a course project. For a squat, for example, you can look at approximate depth, back angle, knee angle, and whether the movement is very asymmetric. It is not as fancy as a trained model, but it is much easier to debug and much easier to present. Fit3D can still be useful, especially for understanding poses and testing ideas, but I would not trust it as the only source. It is much cleaner than real phone footage. Record some videos yourselves very early, even if they are ugly. Those clips will show you what actually breaks. DO NOT spend most of the month training a model. First make the live demo reliable. If the real-time tracking is stable, even a simple rule-based feedback system will look decent. If tracking is unstable, a classifier will not save it.