FGO-PMB: A Factor Graph Optimized Poisson Multi-Bernoulli Filter for Accurate Online 3D Multi-Object Tracking

Area of Science:

• Robotics
• Computer Vision
• Probabilistic Robotics

Background:

• LiDAR-based autonomous systems require reliable perception.
• LiDAR data presents challenges like sparsity, occlusion, and noise, impacting tracking stability.
• Existing methods struggle with uncertainty and instability in 3D multi-object tracking (3D MOT).

Purpose of the Study:

• To develop a unified probabilistic framework for robust 3D MOT using LiDAR.
• To address uncertainty and instability issues in LiDAR-based object tracking.
• To improve the reliability of perception in autonomous systems.

Main Methods:

• Proposed FGO-PMB: a framework integrating Poisson Multi-Bernoulli (PMB) filter (Random Finite Set theory) with Factor Graph Optimization (FGO).
• Formulated object states, existence probabilities, and association weights as optimizable variables in a factor graph.
• Defined four factors (state transition, observation, existence, association consistency) to encode spatio-temporal constraints.

Main Results:

• Achieved temporally consistent and uncertainty-aware estimation across LiDAR scans.
• Demonstrated competitive 3D MOT accuracy on KITTI and nuScenes datasets.
• Maintained real-time performance.

Conclusions:

• FGO-PMB offers a robust solution for LiDAR-based 3D MOT by unifying RFS uncertainty modeling with FGO global optimization.
• The framework effectively handles LiDAR data challenges, enhancing perception for autonomous systems.
• The method provides accurate and stable object tracking in real-time.