Huang Kemiao ; Chen Yinqi ; Zhang Meiying ; Hao Qi

Abstract

Compared with real-time multi-object tracking (MOT), offline multi-objecttracking (OMOT) has the advantages to perform 2D-3D detection fusion, erroneouslink correction, and full track optimization but has to deal with thechallenges from bounding box misalignment and track evaluation, editing, andrefinement. This paper proposes "BiTrack", a 3D OMOT framework that includesmodules of 2D-3D detection fusion, initial trajectory generation, andbidirectional trajectory re-optimization to achieve optimal tracking resultsfrom camera-LiDAR data. The novelty of this paper includes threefold: (1)development of a point-level object registration technique that employs adensity-based similarity metric to achieve accurate fusion of 2D-3D detectionresults; (2) development of a set of data association and track managementskills that utilizes a vertex-based similarity metric as well as false alarmrejection and track recovery mechanisms to generate reliable bidirectionalobject trajectories; (3) development of a trajectory re-optimization schemethat re-organizes track fragments of different fidelities in a greedy fashion,as well as refines each trajectory with completion and smoothing techniques.The experiment results on the KITTI dataset demonstrate that BiTrack achievesthe state-of-the-art performance for 3D OMOT tasks in terms of accuracy andefficiency.