Trajectory-Centric 3D Multi-Object Tracking

Published:

2023 - 2025
Zhejiang University

This research experience focuses on robust 3D multi-object tracking for autonomous driving, especially identity consistency, pose-aware association, and trajectory continuity under long range observations, occlusion, and ego-motion.

  • Developed pose-aware association strategies for online 3D multi-object tracking, improving identity consistency through ego-motion-aware spatial reasoning.
  • Achieved leaderboard-leading performance on the KITTI 3D Multi-Object Tracking benchmark in both online and offline settings, ranking first among submitted methods at the time of evaluation.
  • Extended online tracking into a trajectory-centric offline optimization framework for evidence recovery, topology repair, gap completion, and kinematic state refinement.

PB-MOT: Pose-aware Association Boosted Online 3D Multi-Object Tracking

Status: accepted by IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.

PB-MOT addresses online 3D MOT for robotic and autonomous driving platforms. The method combines ego-motion-compensated state estimation with a rotated ellipse association strategy, aiming to preserve geometric precision and motion robustness while keeping the tracker computationally efficient.

Links: Publication · Video Demo · KITTI Benchmark

PB-MOT++: Offline Trajectory Refinement for 3D Multi-Object Tracking

Status: manuscript under review.

PB-MOT++ extends the pose-aware tracking line toward trajectory-centric offline optimization. Instead of repeatedly operating on frame-wise detections, it optimizes trajectories through boundary recovery, topology repair, gap completion, and kinematic refinement. This direction is intended to improve long-range trajectory consistency with limited computational overhead.

Links: Manuscript Record · Video Demo · KITTI Benchmark

Video Demonstrations

PB-MOT

Open full video

PB-MOT++

Open full video