Gesture Recognition in Tennis Biomechanics

Abstract

The purpose of this study is to create a gesture recognition system that interprets motion capture data of a tennis player to determine which biomechanical aspects of a tennis swing best correlate to a swing efficacy. For our learning set this work aimed to record 50 tennis athletes of similar competency with the Microsoft Kinect performing standard tennis swings in the presence of different targets. With the acquired data we extracted biomechanical features that hypothetically correlated to ball trajectory using proper technique and tested them as sequential inputs to our designed classifiers. This work implements deep learning algorithms as variable-length sequence classifiers, recurrent neural networks (RNN), to predict tennis ball trajectory. In attempt to learn temporal dependencies within a tennis swing, we implemented gate-augmented RNNs. This study compared the RNN to two gated models; gated recurrent units (GRU), and long short-term memory (LSTM) units. We observed similar classification performance across models while the gated-methods reached convergence twice as fast as the baseline RNN. The results displayed 1.2 entropy loss and 50 % classification accuracy indicating that the hypothesized biomechanical features were loosely correlated to swing efficacy or that they were not accurately depicted by the sensor