Robust online multi-object tracking with conditional diffusion motion hypotheses and time-aware contrastive prototypes.

Online multi-object tracking (MOT) is typically implemented as tracking-by-detection: a motion prior propagates each track to the current frame, and appearance cues drive data association. Despite strong detectors and association heuristics, two practical failure modes persist in crowded, interaction-heavy videos: (i) deterministic, unimodal propagation over-commits under non-linear motion and short-term ambiguity, causing overlap-based gating to discard correct matches; and (ii) appearance embeddings drift over time, so stale historical features can dominate similarity scores after long gaps. This paper proposes DiffuTrack, a generative online MOT framework that addresses both issues while retaining the standard predict-associate-update loop. The Motion Diffusion Module (MDM) replaces point motion propagation with a conditional diffusion generator over normalized bounding-box states, producing distributional hypotheses via accelerated DDIM sampling. To stabilize identity association, Time-Aware Prototype Contrastive Learning (TPCL) maintains temporally decayed identity prototypes and trains embeddings with a time-weighted contrastive objective that down-weights stale positives. Experiments on MOT17/MOT20 and DanceTrack under a shared-detection pedestrian-tracking protocol show consistent gains in association-centric metrics, with the largest improvements concentrated on highly non-linear trajectories and occlusion-heavy sequences. Diagnostic uncertainty-coverage analyses further indicate that diffusion-based hypotheses retain a wider region of support than linear-Gaussian propagation, supporting probabilistic motion generation as a practical alternative to deterministic tracking priors in the standard online MOT loop.