A Multi-Object Tracking Method Based on Adaptive Motion Association for Autonomous Vehicle

Although recent multi-object tracking (MOT) methods have achieved considerable progress, tracking from a moving vehicle remains challenging because ego-motion, scale variation, occlusion, and uncertain detections degrade motion prediction and data association. To address these issues, we propose an adaptive motion association framework consisting of a task-aligned enhancement head (TAE-Head), adaptive motion estimation (AME), and extended shape-constrained IoU (ES-IoU). TAE-Head improves the consistency between classification and localization features. AME adjusts process and measurement covariances, estimates a confidence-weighted common displacement from provisional matches, corrects predicted positions, and recomputes association before the final state update. ES-IoU increases geometric tolerance while suppressing shape-incompatible matches. On BDD100K, the proposed framework achieves 56.7% HOTA and 65.0% IDF1, exceeding the baseline by 9.0 and 7.7 percentage points, respectively, while operating at 32.5 FPS. On MOT17, it achieves 68.2% HOTA and 80.5% IDF1.