LSMOTR: An End-to-End Multi-Object Tracking Model with Elegant Use of Long-Short Memory

The multi-object tracking (MOT) tasks require algorithms to maintain the identity of objects over a long time period, making it necessary to maintain long-term features that describe the trajectories. However, current algorithms that utilize long-term memory features either lack expressive power in representing trajectory features or require significant computational resources. Therefore, we propose a novel memory generation and fusion network that enables an end-to-end multi-object tracking model, called LSMOTR. Our model consists of an output generator for producing tracking results and a track query updater. Compared to previous multi-object trackers, our approach incorporates more information from the temporal sequence of video frames. Benefited from such merits, our approach enhances tracking performance in cases of severe object deformations and occlusions while maintaining relatively low computational resource requirements. Through experiments, we achieve highly competitive results on the challenging multi-object tracking dataset, DanceTrack.