MixFEAN: Enhancing Multi-Object Tracking for Intelligent Sports Analysis through Conditional Feature Mixing and Dynamic Re-weighting
Zhihao Zhang, Wan Ahmad Munsif Bin Wan Pa, Chen Zhang, Nur Shakila Mazalan, Wenyue Liu · 2024
Multi-object tracking (MOT) technology plays a key role in intelligent sports event analysis, and is essential for automating athlete data collection, assisting referee decision-making, and enhancing spectator experience. However, existing MOT systems often face challenges such as inconsistent detection accuracy and recognition difficulties, which stem from problems caused by the fast and complex context of the sport. To address these challenges, we propose an innovative approach called Mixed Feature Enhanced Attention Network (MixFEAN), which aims to improve the performance of multi-target tracking systems in complex scenarios through an integrated strategy of data enhancement and feature enhancement. Specifically, we introduce a feature-level conditional mixed data enhancement technique, FcMixup, which mixes the most similar feature representations during training to enhance the model’s adaptability to different motion behaviors and environmental changes. In addition, we design a novel Feature Decoupling Re-weighting Module (FDRM) that dynamically adjusts the weights of the feature maps so that the model is more focused on the identification and tracking of key targets, thus effectively improving the detection accuracy and enhancing the anti-jamming capability. The uniqueness of MixFEAN lies not only in the diversified data, but also in the optimization of the feature extraction through the noticing mechanism, thus enhancing the overall performance of the system while Maintaining Efficiency. We conducted extensive experiments on public datasets such as LSP and SportsMOT. The results show that MixFEAN improves the accuracy by $1.1 \%$ on the LSP dataset and significantly improves it on the SportsMOT dataset. These findings not only demonstrate the feasibility and effectiveness of the MixFEAN method in real-world scenarios, but also provide strong technical support for the further development of intelligent sports monitoring systems.