TrackV2X: Cross-Frame Data Correlation for V2X 3D Multi-Object Tracking

3D multi-object tracking (MOT) is a core perception task in autonomous driving. Vehicle-to-Everything (V2X) communication enhances the perception capabilities of autonomous systems by fusing data from various sensors on both vehicles and infrastructure. However, research on tracking methods for V2X scenarios remains limited. This paper presents TrackV2X, a 3D multi-object tracking method specifically designed for V2X environments. The method follows the tracking-by-detection paradigm, where detections are first obtained through an object detector, and then the shape and spatio-temporal affinities between detection boxes in consecutive frames are learned. In this context, affinity serves as a probabilistic matching approach. By learning the affinities between detection boxes, the method enables the correlation of detections across consecutive frames. This results in robust cross-frame data correlation, which eliminates false positives, recovers false negatives, and accomplishes the sequence tracking task. Experimental results on the V2X-Seq dataset demonstrate the accurate tracking performance of the proposed method.