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

## TL;DR

This paper introduces FGO-PMB, a new method for tracking multiple objects in 3D using LiDAR data, which improves accuracy and stability in autonomous systems.

## Contribution

The novel contribution is a unified probabilistic framework combining PMB filters and FGO for robust 3D multi-object tracking.

## Key findings

- FGO-PMB achieves temporally consistent and uncertainty-aware object tracking.
- The method performs competitively on KITTI and nuScenes datasets while maintaining real-time performance.

## Abstract

Three-dimensional multi-object tracking (3D MOT) plays a vital role in enabling reliable perception for LiDAR-based autonomous systems. However, LiDAR measurements often exhibit sparsity, occlusion, and sensor noise that lead to uncertainty and instability in downstream tracking. To address these challenges, we propose FGO-PMB, a unified probabilistic framework that integrates the Poisson Multi-Bernoulli (PMB) filter from Random Finite Set (RFS) theory with Factor Graph Optimization (FGO) for robust LiDAR-based object tracking. In the proposed framework, object states, existence probabilities, and association weights are jointly formulated as optimizable variables within a factor graph. Four factors, including state transition, observation, existence, and association consistency, are formulated to uniformly encode the spatio-temporal constraints among these variables. By unifying the uncertainty modeling capability of RFS with the global optimization strength of FGO, the proposed framework achieves temporally consistent and uncertainty-aware estimation across continuous LiDAR scans. Experiments on KITTI and nuScenes indicate that the proposed method achieves competitive 3D MOT accuracy while maintaining real-time performance.