FusionAction Net: a real-time sports action quality evaluation algorithm based on GCN and improved DTW

To meet the technical requirements for the automatic, real-time quality evaluation of sports actions, this paper proposes a real-time evaluation algorithm that combines a graph convolutional network with improved dynamic time warping. To address the difficulty traditional methods face in accounting for the spatio-temporal characteristics of actions and the fine-grained measurement of execution timing, this study designs an innovative model, FusionAction Net. The model first extracts deep semantic features of the skeleton sequence using an attention-enhanced spatiotemporal Graph Convolutional Network module; then, an enhanced Dynamic Time Warping module with semantic constraints is designed to measure the timing execution deviation of actions in detail. End-to-end scoring output is realized through feature fusion. To verify the algorithm’s effectiveness, this study built a high-quality skeleton dataset comprising five fundamental sports actions and 1250 samples. The experimental results show that FusionAction Net achieves a macro-averaged accuracy of 93.9% in action recognition, with a correlation coefficient of 0.92 between its output scores and expert evaluations, and an average absolute error of 0.85 points. In a standard hardware environment, the model achieves a frame-level spatial feature extraction speed of about 12 milliseconds per frame and an end-to-end processing time of 122 milliseconds per sequence, fully meeting the requirements for real-time evaluation. This study provides a high-precision and practical solution for intelligent sports training.