MoSA-Det: motion state adaptive object detection for sports videos

Object detection in sports videos serves as a fundamental task for applications such as intelligent broadcasting, tactical analysis, and athlete tracking. Existing methods face two critical challenges when processing sports scenarios: (1) Motion blur-induced feature degradation, where fast-moving objects generate severe motion blur that leads to indistinct boundaries, texture loss, and significantly reduced feature response intensity, severely weakening the discriminative capability of detectors; (2) Temporal aggregation failure, where existing temporal methods assume small inter-frame displacements that enable effective alignment, yet fast-moving objects often exhibit excessive inter-frame displacement that causes alignment failure, making multi-frame aggregation introduce noise and degrade detection accuracy. To address these challenges, we propose MoSA-Det, a framework whose core idea is to leverage motion states as regulatory signals for detection strategies, achieving joint adaptive optimization of feature extraction and temporal fusion. The framework comprises two core modules: the Motion-Aware Adaptive Feature Module (MAAF) and the State-Guided Temporal Aggregation Module (SGTA). Specifically, MAAF constructs a lightweight motion state estimator through inter-frame feature differencing and local correlation analysis to generate fine-grained motion state priors, and employs a state-conditioned dynamic convolution mechanism to perform weighted fusion of multi-scale receptive field features, introducing deformable convolution for adaptive spatial sampling in high-speed motion regions to effectively alleviate feature degradation caused by motion blur. SGTA embeds motion state priors into the temporal feature fusion process, achieving selective aggregation of inter-frame features through state-aware adaptive weights, ensuring that static regions fully exploit multi-frame information to enhance feature stability while enabling fast-moving regions to avoid noise interference from misaligned frames. Experiments on the SoccerNet-Tracking and SportsMOT datasets demonstrate that MoSA-Det improves mAP@0.5 by 1.7% and 1.6%, and mAP@0.75 by 2.6% and 1.8% over the strongest baselines, respectively, validating the effectiveness of the motion state adaptive strategy.