Tags from personal dataset were analyzed with MediaPipe Holistic for body posture, hand and face signatures and concatenated for LSTM model.
The model is made up of 3 LSTM layers and 3 Dense layers.
They have a precision of 90% for highly similar actions and 78% for dissimilar actions.