An Adaptive Motion Tracking Method for Sports Analytics

Multi-object tracking presents a difficult task, especially in the dynamic environment of sports analytics, where the movement of sports players is continuous. Players on the same team have similar appearances, making continuous player tracking challenging. Furthermore, occlusions by players may prevent a given player's detection in numerous frames. In this paper, we propose a novel tracking method, Adaptive Motion Track, which enables motion-estimation modelling of sports players. Firstly, we replace the Kalman Filter, a traditional motion-estimation method for bounding box prediction, with a Multi-Layer Perceptron model for bounding box estimation. Secondly, we propose a novel two-stage tracking approach based on the established BYTE structure with each stage using a different motion-estimation distance metric: intersection over union for the first, and Euclidean distance for the second. Additionally, we incorporate appearance-based features for object identification within the proposed tracking structure. Our method follows the Tracking-by-Detection paradigm, enabling flexibility as detection algorithms can be changed, essentially future-proofing the tracking method. The proposed method outperforms current methods, achieving SOTA performance when tracking players using the TeamTrack-Soccer Side Validation dataset.