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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Online multi-object tracking (MOT) commonly uses tracking-by-detection, facing challenges with non-linear motion and appearance drift.
• Existing methods struggle in crowded, interaction-heavy videos due to deterministic motion propagation and stale appearance features.

Purpose of the Study:

• To introduce DiffuTrack, a generative online MOT framework addressing limitations in motion prediction and identity association.
• To improve tracking robustness in challenging scenarios like occlusions and non-linear movements.

Main Methods:

• Implemented a Motion Diffusion Module (MDM) for probabilistic motion hypothesis generation using conditional diffusion.
• Introduced Time-Aware Prototype Contrastive Learning (TPCL) to stabilize identity association by managing appearance embedding drift.
• Retained the standard predict-associate-update loop within the generative framework.

Main Results:

• DiffuTrack demonstrated consistent gains in association-centric metrics on MOT17/MOT20 and DanceTrack datasets.
• Significant improvements were observed for non-linear trajectories and occlusion-heavy sequences.
• Diffusion-based hypotheses showed wider support regions compared to traditional linear-Gaussian propagation.

Conclusions:

• DiffuTrack offers a practical generative approach to online MOT, overcoming limitations of deterministic priors.
• Probabilistic motion generation and temporally aware appearance learning enhance tracking performance in complex scenarios.
• The framework provides a viable alternative for robust online multi-object tracking.