I’m very new to machine learning, and am currently trying to write a program that detects correct and incorrect sitting posture using Tensorflow, Keras, and Movenet. At the moment, I have good and bad posture folders containing the respective posture in train, valid, and test folders. I have a couple ideas on ways I would go about coding the model, but I’m unsure about which would actually work/be the most appropriate:
-
Keep good/bad posture folders, build and train a CNN using the data, and then build/train the fine-tuned neural network to classify good and bad posture
-
Only keep good posture images, use slope to calculate correct positions of certain body parts, no CNN building. Only problem I can think of with this is what if a certain body part was obstructing another body part? For example, I have a picture of correct posture with my arms on the table and positioned for typing, but I also have another picture of correct posture with my arms to my side, where the forearm down isn’t visible. or with incorrect posture, what if someone was leaning to the side and their face was resting on their hand, so some body parts aren’t visible? How would cases like that be handled by me or Movenet?
-
A completely different approach altogether
If anyone could help point me in the right direction, I’d really appreciate it! Thanks in advance!