A Study on Recognizing and Analyzing Aerobics Movements by AI based Video Feature Extraction

The identification and investigation of various aerobics movements holds immense practical importance in the field of aerobics for educational purposes. The study initially obtained skeletal data from aerobics videos using the Azure Kinect DK camera. Subsequently, several variables, including coordinates, velocity, acceleration, and vectors of joint locations, were extracted from the collected data. The features were combined and utilized as input for the algorithm that recognizes and analyses movement. The AI-based SA-BiLSTM algorithm was created by combining bidirectional long short-term memory (LSTM) with a self-attention mechanism (SA). The NTU RGB+D dataset and a self-built aerobics dataset were utilized for conducting experiments. The results demonstrated that the fused feature outperformed individual features in terms of recognition and analysis performance. The SA-BiLSTM algorithm, which is based on AI, attained an accuracy of 91.27% in CS (Computer Science) and 96.58% in CV (Computer Vision), respectively. It outperformed methods such as ST-GCN. The AI-based SA-BiLSTM algorithm achieved an average accuracy of 94.83% in recognizing and analyzing aerobics movements, surpassing the LSTM and BiLSTM techniques which achieved 81.86% and 90.81% respectively. The findings presented here showcase the dependability of the AI-powered SA-BiLSTM algorithm in recognizing and analyzing actions, thereby making it suitable for real- world applications.