Athlete Training Movement Recognition using Long Short-Term Memory Model with Adaptive Step Size Based Crow Search Algorithm

With the rapid growth of recognition technology in images, human action recognition has become increased rapidly, particularly in the context of Athlete Training Movement Recognition (ATMR). This research addresses the challenge of recognizing athlete behaviors in sports teaching videos concentration on the depth frames. In order to improve the recognition accuracy for various athletic behaviors, the research involves extracting and selecting meaningful features from the video frames. The Adaptive Step Size-Based Crow Search Algorithm (ACSA) is employed to select the relevant features from the 4096 extracted VGG-16 deep features and these features are then used to trains a Long Short-Term Memory (LSTM) networks. LSTM network is adopted to differentiate the action classes and evaluate with performance metrices such as accuracy, Positive Predictive Value (PPV), recall, specificity and F-measure. When compared with other models like Recurrent Neural Networks (RNN), Gated Recurrent Units (GRU), and Convolutional Neural Network (CNN), proposed model achieved an impressive accuracy of 95.54% in athlete training movement recognition.