Answers are grounded in this paper and cite their sources.
Multi-object tracking (MOT) is a vital task in computer vision, involving the continuous tracking and stable identification (ID) of multiple objects across video sequences. Existing methods often focus heavily on identity matching during association, neglecting the quality of trajectories, which affects object tracking performance. This article proposes a trajectory-age optimization approach for MOT that effectively improves its performance. First, we proposed TAE and TG to adaptively extend trajectory lifespans through IoU constraints and position awareness, allowing long-lost objects to be re-associated with their original trajectories instead of being initialized as new tracks, while suppressing erroneous matches. Second, by incorporating the track loss duration, our method enhances the observation-centric re-update mechanism and an adaptive Kalman filter with a dynamic noise scale, thereby reducing ID switches. Finally, our method combines Gaussian process-based interpolation with wavelet-thresholding denoising, leveraging the optimization of global smoothing and local denoising to improve trajectory smoothness and authenticity. Extensive experiments on the MOT17, MOT20, and DanceTrack standard datasets demonstrate that our method is competitive with the current state-of-the-art trackers and offers more robust ID maintenance across various challenging tracking scenarios.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: