DSSO: A High-precision Action Recognition Model for Aerobics based on Dynamic Segmentation and Sequence Optimization

With the development of artificial intelligence and computer vision technologies, the application of these advanced technologies to the movement analysis of athletes has become a research focus in the field of sports science. Especially in aerobics, a sports activity that requires a high degree of coordination and technical precision, the application of movement recognition technology can significantly enhance training effects and athletic performance. However, due to the rapidity of aerobics movements and the interference of complex backgrounds, existing action recognition systems have obvious limitations in capturing and analyzing continuous action sequences, such as motion occlusion and insufficient detail capture, which seriously affect the application of the technology. In this study, we propose an integrated action recognition system that contains three modules, aiming to solve the above problems. First, an innovative foreground separation technique is used to effectively extract the athlete’s image from a complex or dynamic background to clearly separate the action subject. Next, a super-resolution technique is introduced to enhance the details of the foreground image, a step that significantly improves the system’s accuracy in recognizing fast and small movements. Finally, the recognition process of the entire aerobics action sequence is optimized by means of time series analysis to effectively deal with the recognition challenges brought by action occlusion and rapid changes. Experimental results on HSiPu2 and UESTC RGB-D Varying-view action databases show that the present model can significantly improve the accuracy compared with existing models.