RobMOT: Robust 3D Multi-Object Tracking by Observational Noise and State Estimation Drift Mitigation on LiDAR PointCloud

This paper addresses limitations in 3D tracking-by-detection methods,particularly in identifying legitimate trajectories and reducing stateestimation drift in Kalman filters. Existing methods often use threshold-basedfiltering for detection scores, which can fail for distant and occludedobjects, leading to false positives. To tackle this, we propose a novel trackvalidity mechanism and multi-stage observational gating process, significantlyreducing ghost tracks and enhancing tracking performance. Our method achieves a$29.47\%$ improvement in Multi-Object Tracking Accuracy (MOTA) on the KITTIvalidation dataset with the Second detector. Additionally, a refined Kalmanfilter term reduces localization noise, improving higher-order trackingaccuracy (HOTA) by $4.8\%$. The online framework, RobMOT, outperformsstate-of-the-art methods across multiple detectors, with HOTA improvements ofup to $3.92\%$ on the KITTI testing dataset and $8.7\%$ on the validationdataset, while achieving low identity switch scores. RobMOT excels inchallenging scenarios, tracking distant objects and prolonged occlusions, witha $1.77\%$ MOTA improvement on the Waymo Open dataset, and operates at aremarkable 3221 FPS on a single CPU, proving its efficiency for real-timemulti-object tracking.