Reproduce “OptiPMB: Enhancing 3D Multi-Object Tracking with Optimized Poisson Multi-Bernoulli Filtering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guanhua Ding, Yuxuan Xia, Runwei Guan, Qinchen Wu, Tao Huang, Weiping Ding, Jinping Sun, Guoqiang Mao
OptiPMB is an optimized Poisson Multi-Bernoulli (PMB) filter designed for 3D multi-object tracking in autonomous driving, integrating adaptive birth, detection probability, and refined pruning. The method achieved state-of-the-art online tracking accuracy on nuScenes and KITTI datasets, notably reducing identity switches and false positives compared to prior model-based approaches.
Key insights
- RFS-based Bayesian filters, when specifically optimized for autonomous driving challenges, can achieve state-of-the-art 3D MOT performance, demonstrating their competitiveness against learning-based or heavily heuristic RV-based methods.
- The Hybrid Adaptive Birth Model (HABM) is identified as the most impactful component, significantly improving track initialization, reducing identity switches, and enhancing overall tracking accuracy by generating reliable newborn objects and suppressing false positives.
- The various adaptive designs and optimization modules in OptiPMB work synergistically, with their combined integration yielding superior overall performance gains compared to individual contributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.