A Cue-Level Fusion Framework for Robust Soccer Player Tracking: Integrating Appearance, Motion, and Geometric Cues for Adaptive Tracklet Linking

Cue-level fusion of complementary information channels—appearance for identity, motion for trajectory, and geometry for spatial consistency—is central to multi-object tracking in soccer broadcast videos, yet existing tracking-by-detection frameworks are validated mainly on pedestrian data. This paper proposes a cue-level fusion framework for soccer player tracking on the SoccerNet SN-Tracking dataset (57 train, 49 test sequences). Fusion is performed at two levels: (i) motion-appearance gating during online association, fusing Mahalanobis distance and cosine distance through gated cascade matching; and (ii) cue-level post-processing, jointly fusing temporal, spatial, and geometric cues for tracklet linking. Our contributions are: (1) a systematic cue-contribution analysis over ten configurations, quantifying independent and interaction effects—including the pure contribution of the OSNet appearance model—and showing that appearance is the dominant contributor (+19.00 HOTA) while the original linking cue suffers from gain saturation on high-quality trackers; (2) a hyperparameter sensitivity analysis identifying soccer-specific best-performing configurations, including motion retention at 25 frames (~1 s at 25 fps), which is fixed by a pre-specified detection-gap rule established on the training split and is significantly better than neighboring values on the test split (20→25: +2.80, t(48) = 5.57, p < 0.001, Cohen’s d = +0.80; 25→30: −3.12, p < 0.001); and (3) a multi-cue veto cascade (MCVC), a training-free and strictly subtractive team-color equivalence-class veto that overcomes saturation by rejecting cross-team links—including those involving outlier entities such as referees and the ball—instead of re-scoring them, achieving gains of +2.22 HOTA and +4.04 AssA over the original AFLink (paired per-sequence test against the no-veto control, p < 0.001 on both splits). The best configuration achieves 64.31% HOTA on the test set, providing a configurable cue-level baseline for soccer video tracking.