Optimal Time-Window Derivation for Human-Activity Recognition Based on Convolutional Neural Networks of Repeated Rehabilitation Motions.