Cooperative 3D Multi-Object Tracking With Cross-Agent Data Association for Connected and Automated Vehicles

Achieving reliable 3D multi-object tracking (MOT) is essential for enabling autonomous driving. With the aid of communication techniques, vehicles could share complementary environmental information, ushering in an innovative break through in perception performance. Nevertheless, most tracking methods focus on single-agent scenarios, which inevitably struggle to cope with temporarily disappearing objects and incapable of leveraging multi-agent cooperation. In this study, we propose a communication-efficient cooperative 3D MOT framework. Specifically, to suppress the occurrence of temporary tracking failures and ID switching, we design a cross agent data association module, which captures the correspondence between objects tracked by multiple agents to compensate missed objects while concealing erroneous tracking for individuals. Moreover, a search area expansion mechanism is introduced to assist the tracker in handling previously missed objects. The extensive experiments on cooperative perception dataset demonstrate that the proposed method is able to reduce both misses and erroneous detections, and it performs favorably against single-agent trackers without imposing a communication burden on other perceptual tasks.