TrajMOT: A Multi-Object Tracking Framework via Trajectory Prediction

Consistency in 3D tracking is critical for vision-based autonomous driving systems. However, most existing multi-object tracking (MOT) methods are confronted with the limitation: detectors underutilize temporal cues, while trackers struggle to capture long-term dependencies. In response to these challenges, we introduce, we propose TrajMOT, a framework that enhances tracking consistency through unified spatiotemporal modeling of object trajectories across multi-view inputs. In detail, our approach adopts a tracking-by-attention framework, where object queries maintain coherent instance representations over time. In the detection module, TrajMOT utilizes bird's-eye-view (BEV) queries to interact with spatial features through spatial cross-attention and integrates historical BEV information using temporal self-attention. The tracking module processes historical information and predicts robust future trajectories by employing cross-frame attention, boosting feature stability and discriminability for identical objects. Our method effectively maintains the identity consistency of objects by integrating trajectory predictions during periods of long-term occlusions or image blurs. Evaluations on the nuScenes dataset demonstrate that our model achieves significant improvements in Average Multi-Object Tracking Accuracy (AMOTA), with an 8.7% increase, and a 15% reduction in ID switches (IDS), compared to prior methods.