This study proposes an enhanced motion adaptive StrongSORT framework for multi-object tracking in sports scenarios, addressing tracking failures caused by rapid athlete motion, frequent occlusions, and complex interactions. First, a probabilistic motion state recognition module based on Bayesian classification and a hidden Markov model is introduced to categorize athlete behavior into five states: stationary, constant speed, acceleration, deceleration, and turning. Second, to address numerical divergence caused by high-speed maneuvers, a numerically stable adaptive Kalman filter is applied, employing the standard Joseph-form covariance update and a bounded adaptive mechanism from control theory, thus ensuring numerical stability under finite-precision arithmetic. Third, a physics-aware occlusion management strategy integrating distance intersection over union (DIoU), complete intersection over union (CIoU), and a graph neural network is proposed to maintain identity consistency in complex interactions. Finally, a quality-aware appearance modeling framework is established, in which motion-state-adaptive momentum updating is used to prevent feature degradation. Experimental results on the SportsMOT dataset show that the MOTA, IDF1, and MT metrics of the proposed method reach 82.9%, 85.3%, and 68.2%, respectively. Compared with the baseline StrongSORT, these metrics improve by 8.3, 8.8, and 13.5 percentage points, respectively, while identity switches are reduced by approximately 39 percentage points. A real-time processing speed exceeding 30 frames per second is maintained, supporting online sports video analysis.