Real-time optical motion capture and technical statistics system for basketball games based on deep learning

In response to the conflict between real-time and accuracy of motion capture in high-dynamic scenes in basketball games, as well as the lack of robustness of multimodal data fusion and three-dimensional trajectory reconstruction in complex interactive scenes, this paper introduces a deep learning-based method to fuse multicamera information through a multiview Transformer to reconstruct the posture of occluded players and reduce tracking errors under occlusion. Combining the optical flow histogram with the spatiotemporal attention mechanism, cross-modal time alignment of visual data and inertial sensor data is achieved to solve the problem of data misalignment in high-speed motion. In this system, optical technology is integrated into motion capture, and this mature technique improves the precision of motion tracking under dynamic occlusion. The introduction of optical flow features enables the system to more sensitively extract detailed motion vectors, further enhancing its robustness in complex game environments. Based on the lightweight 3D convolutional network optimized by MobileNet, the motion recognition accuracy reaches 88.6%-92.1% in different occlusion scenarios. Experimental results show that the system is superior to traditional methods in terms of motion recognition accuracy, data synchronization and inference delay. In occlusion scenarios, the accuracy is improved by 7.59%-35.89%. It can generate more than 100 tactical indicators in real time, providing data support for real-time analysis and tactical decision-making in basketball games.