Adaptive Searching Range-Based Data Association for Multi-Object Tracking With Multi-Information Fusion

The goal of multi-object tracking (MOT) is to estimate the location of objects and maintain their identities consistently to yield their individual trajectories. It has become a trend to fuse multi-sensor information to achieve 3D MOT, since it can leverage the advantages of different sensors to enhance tracking performance. However, it is a challenging work due to the necessity of fusing features with diverse attributes and wrong association caused by significant noise. In this paper, we propose adaptive weight parameter-based multi-feature fusion to create affinity function, alongside adaptive setting of data association searching range, aiming to augment camera-Lidar information fusion-based MOT framework. First, detected results from these two sensors are divided into three categories. Then, to fully utilize both motion and appearance information, adaptive weight parameter setting is proposed to embed appearance information into motion information, forming the basis for creating an affinity function for data association. Furthermore, to mitigate wrong association caused by object temporary occlusion or out-of-view, a method for adaptively adjusting the association searching area is introduced based on the number of frames in which tracking trajectories disappear. Finally, to prevent appearance information pollution caused by significant noise, a confidence score-based tracking trajectory appearance feature updating strategy is explored. The experiment results on KITTI and nuScenes MOT benchmark show remarkable performance improvement over other state-of-the-art MOT methods and demonstrate the effect of our designed modules.