Pirouette Recognition LSTM

I worked from a public IMU dataset of about 85 labeled turns across six dancers, backing out joint torques from inertia estimates to run alongside the raw time series. I trained an LSTM to classify turn type from the combined signal (see image left).

It reached over 95% accuracy, though with a dataset that small the number says more about fitting six dancers than about generalizing to new ones. Written up under Professor Phuong Truong's guidance.