关键词: 在线多目标跟踪, 自适应融合, Mamba, 关联不确定性, 状态空间模型, 运动预测, 遮挡处理

Abstract: To improve identity preservation in dense overhead fisheye scenes affected by radial distortion and frequent occlusion, a dual-stream Mamba online multi-object tracking method addressing association uncertainty was proposed. Motion-Mamba and Appearance-Mamba were used to model trajectory-level motion and appearance temporal priors, respectively. Historical and predicted geometric consistency were combined to construct the Association Ambiguity Interval (AAI), which partitioned candidate pairs into confident, ambiguous, and non-association regions. GrayFusion was activated only for AAI candidates and adaptively fused motion and appearance cues according to their confidence difference and appearance timeliness. On the Fisheye Overhead Multi-person Dense (FOMD) test set, the proposed method achieves an Identification F1 Score (IDF1) of 86.72% and a Multiple Object Tracking Accuracy (MOTA) of 82.73%, with 7 Identity Switches (IDSW). Compared with the Intersection over Union (IoU) baseline, IDF1 and MOTA increase by 10.89 and 3.55 percentage points, respectively, and IDSW decreases by 26. Compared with BoostTrack++ and ClarityTrack, IDF1/MOTA increase by 0.57/0.34 and 2.77/2.17 percentage points, respectively, while IDSW decreases by 5 and 12. Experimental results show that the proposed method reduces identity switches under dense occlusion and nonlinear motion while maintaining high overall tracking accuracy, and provides differentiated handling of candidate-level association uncertainty.

Key words: online multi-object tracking, adaptive fusion, Mamba, association uncertainty, state space model, motion prediction, occlusion handling