Research on human motion capture in sports videos based on multiscale transformer

With the continuous development of computer technology, the impact of digitalization on the sports industry is becoming increasingly profound. However, the volume of sports video data is large and the complexity is high, making it difficult to capture and identify human movements from the relevant video data. Therefore, this study proposes a sports video human motion capture model based on multi-scale Transformer. This model achieves multi-scale feature alignment and fusion in the encoder-decoder structure by introducing a multi-scale spatio-temporal feature extraction mechanism and a separation pooling module, and builds an intelligent sports motion capture system, reducing the operational difficulty of the model. Experiments show that the improved model achieves an average capture accuracy rate of up to 97.6% in multiple sports action recognition tasks, which is 2.7% higher than that of the second-best spatio-temporal graph convolutional network model. Moreover, in the test set, the capture accuracy rate only decreases by 0.8%, demonstrating good generalization ability and robustness. The improved model's accuracy rate for capturing all movements is above 90%, and the accuracy rates for capturing short and long movements are 90.4% and 91.2% respectively. It can be concluded from this that the model proposed by the research institute can effectively improve the accuracy of human motion capture and promote the refined research of sports science.