3D Multi-Object Tracking based on Two-Stage Data Association for Collaborative Perception Scenarios

Abstract

This paper proposes a 3D multi-object tracker suitable for collaborative perception scenarios. Our tracker aims to associate detection candidates obtained from the ego-vehicle after performing collaborative perception over time. It considers the temporal asynchronous information exchanged among connected vehicles and focuses on dealing with objects that fail to track due to missed detection. To achieve this, we propose a two-stage data association module with a supplementary mechanism. It adapts the association strategy to track objects according to their state robustly. Specifically, the first stage works on most general objects. The second stage aims to associate spatiotemporal asynchronous detection candidates or tracked objects consecutively missed multiple times. A supplementary mechanism is applied to temporarily missed objects by the detector. We conduct experiments on the DAIR-V2X dataset and use the detection candidates generated by a collaborative detection module. Experimental results demonstrate that the proposed method outperforms baselines in tracking performance while achieving comparable speed.