Sports video analysis is one of the most popular of the many sports applications based on artificial intelligence technology. However, traditional sports video analysis methods often face problems such as insufficient real-time performance, low accuracy in small target recognition, and resource limitations. To address these issues, this study combines the advantages of the deep Simple Online and Realtime Tracking (SORT) and YOLOv8 algorithms to propose a novel target tracking method for complex sports scenes. The method optimizes the YOLOv8 network structure by replacing ordinary convolutions in the detection module’s backbone with separable point-wise dynamic convolution, improving the flexibility and efficiency of feature extraction. Regarding the loss function, the Modified Panoptic Intersection over Union (MPDIOU) loss function is adopted instead of the IOU loss function in YOLOv8, improving the model’s detection accuracy for small targets. In the Deep SORT tracking module, a hybrid attention mechanism (including positional attention and channel attention) is introduced, and the robustness of tracking is significantly enhanced by combining motion prediction and feature matching. The experimental results indicated that each module component proposed by the research had a positive impact on the final model and could effectively raise the accuracy of the model in athlete motion tracking. The mean squared error and mean absolute error were controlled within the ranges of 0.4–0.7 and 0.2–0.6, respectively. The accuracy of the research method reached a max of 97.57%, the recall rate reached a max of 97.64%, and the F1 score reached a max of 97.35%, while maintaining a low average detection time of only 2.31 s. Even in the presence of occlusion and small targets, the research method still demonstrated good performance. From this, the research method has significant advantages in target tracking tasks in complex sports scenes, which can raise the real-time accuracy of video analysis.
Lin et al. (Sun,) studied this question.