An intelligent motion trajectory tracking approach for physical education using multi-target recognition

Against the digital transformation of physical education, this study addresses the limitation of traditional subjective observation methods in dynamic multi-target tracking by proposing an intelligent motion trajectory tracking algorithm based on multi-target recognition. Distinguished from existing multi-attention and tracking methods, the algorithm’s core innovations are threefold: first, a Channel Attention Mechanism based on Multi-Scale Convolution (CAM-MSC) is designed via InceptionV1 to enhance keypoint recognition for small-scale and occluded targets; second, an IM-DenseNet is constructed by embedding an upsampling module and CAM-MSC into DenseNet, which reduces parameters by 30% through depthwise separable convolution while boosting feature reuse; third, an IM-DenseNet-DeepSort model is proposed by fusing GhostNet, FPN and IM-DenseNet to optimize DeepSort, effectively reducing occlusion-induced ID-switching errors. Experiments conducted on the MPI (Human Pose Dataset) and Microsoft Common Objects in Context (MSCOCO) show that IM-DenseNet outperforms ResNet and Res-Attention-Net by 1.78% and 0.98% in pose estimation accuracy with fewer parameters; IM-DenseNet-DeepSort achieves over 96% accuracy for shoulder/knee keypoints and 97.89% for head-trajectory recognition on MSCOCO. Ablation studies quantify that CAM-MSC, IM-DenseNet and the optimized DeepSort improve accuracy by 1.10%, 0.22% and 1.60% respectively. With an acceptable inference time for physical education scenarios, the model promotes the integration of computer vision and sports science, laying a technical foundation for digital and precision-oriented physical education.