Uncertainty-Guided Triple-Memory Fusion for Robust Multi-Object Tracking

<title>Abstract</title> <p>Multi-object tracking in complex dynamic scenes remains challenging due to non-linear motion, frequent occlusion, and appearance variation. Traditional methods rely on fixed motion priors and fail to model uncertainty or fuse multi-modal cues adaptively. This work presents an uncertainty-guided triple-memory fusion framework for robust multi-object tracking. The method constructs motion memory via probabilistic Mamba to quantify prediction uncertainty, along with appearance memory and category memory to maintain stable visual and semantic representations. A dynamic fusion mechanism adjusts cue weights based on motion uncertainty and trajectory state to suppress error accumulation. Here we show that the proposed method achieves HOTA scores of 58.2%, 74.8%, and 48.5% on VisDrone, DanceTrack, and SportsMOT datasets, respectively, outperforming several state-of-the-art approaches. This framework improves identity preservation and tracking stability in dense and high-dynamic scenarios, suitable for real-world surveillance and autonomous perception applications.</p>