Machine Learning Based Pose Detection Algorithm Applied to Volleyball Training

Abstract

This work develops a machine learning-based stance identification system and applies it to volleyball training in order to improve the dynamic analysis of volleyball players' stances while dunking and serving. In this paper, a long short-term memory (LSTM) network is presented as a solution to the issue that more distant historical signals cannot be communicated to the present under the network topology of a recurrent neural network (RNN). Further, for the human behavior action also existing as a time series of long-term dependence problems and the problem of real-time detection caused by the use of the traditional sliding window algorithm to collect data, the LSTM is extended to apply to human posture detection, and the LSTM-based human posture detection method is proposed. In this paper, real-time volleyball training data is captured by HD cameras and manually annotated. The performance of the proposed algorithm is tested on a dataset of 5000 manually annotated volleyball player poses. The proposed method has a high detection accuracy and a short feature extraction time.