Enhancing Multi-Object Tracking in Roadside Perception Through Trajectory Smoothing and Dynamic Variance Adjustment

This study addresses the challenges of multi-object tracking (MOT) in vehicle-infrastructure cooperative systems, which has varying densities of LiDAR points and detection accuracy across different regions while requiring accurate tracking results in all regions. Conventional MOT methods in autonomous driving primarily focus on how to match detected objects with tracked targets and how to optimize the motion model to improve the accuracy of tracking results. Unlike these methods, this paper proposes a novel optimization approach for correcting observation values and adjusting observation variance dynamically to enhance tracking accuracy. The key innovation of this paper lies in smoothing historical trajectories points and calculating the variance of the distance between recent observation values and smoothed historical trajectory to correct significant errors in observation values and adjust observation variance dynamically. Experiments conducted in cooperative vehicle-road scenarios demonstrate significant improvements in tracking accuracy for position, speed, and orientation. The proposed methods reduced average position error from 1.34 m to 0.26 m, speed error from 1.02 m/s to 0.19 m/s, and heading error from 2.29° to 0.89°. These results highlight the effectiveness of the proposed methods in improving MOT performance in vehicle-infrastructure cooperative environments.