Online Multi-Object Tracking Using LiDAR

Abstract

Multi-object tracking (MOT), often referred to as multi-target tracking (MTT), plays a pivotal role in applications such as connected and autonomous vehicles (CAVs) and autonomy through infrastructure (ATI). Positioned as a mid-level task within these systems, MOT serves as the foundation for subsequent applications, encompassing environmental comprehension, behavior analysis, and intelligent decision-making. The principal objectives of MOT include the detection and precise localization of multiple objects within a given scenario, the distinction of their identities, and the generation of highly accurate tracking trajectories.

To augment the perceptual abilities of autonomous agents operating in intricate environments, we present a novel online 3D MOT approach grounded in point cloud data. This method not only enhances prediction accuracy but also introduces a novel data association mechanism.

The method employs a constant acceleration model and incorporates orientation angle variation. Simultaneously, the offsets resulting from the motion of the probe agent are nullified through the utilization of the calibration matrix. After undergoing Kalman filter smoothing, this method can effectively and accurately estimate the future states (position, orientation, velocity, acceleration, etc.) of target objects, mitigating the issue of directional oscillation in the object detection stage. To address the bidirectional pairing problem between predicted and candidate targets in complex scenarios, we introduce a novel spatio-temporal feature-based data association model. This model leverages a dynamic confidence threshold to handle tracking of temporarily occluded objects.

Through extensive evaluations on the KITTI dataset, our method surpasses state-of-the-art methods. The method's performance is further validated on the nuScenes dataset, confirming the robustness and effectiveness of our proposed method.