OptiPMB: Enhancing 3D Multi-Object Tracking with Optimized Poisson Multi-Bernoulli Filtering

Accurate 3D multi-object tracking (MOT) is crucial for autonomous driving, as it enables robust perception, navigation, and planning in complex environments. While deep learning-based solutions have demonstrated impressive 3D MOT performance, model-based approaches remain appealing for their simplicity, interpretability, and data efficiency. Conventional model-based trackers typically rely on random vector-based Bayesian filters within the tracking-by-detection (TBD) framework but face limitations due to heuristic data association and track management schemes. In contrast, random finite set (RFS)-based Bayesian filtering handles object birth, survival, and death in a theoretically sound manner, facilitating interpretability and parameter tuning. In this paper, we present OptiPMB, a novel RFS-based 3D MOT method that employs an optimized Poisson multi-Bernoulli (PMB) filter while incorporating several key innovative designs within the TBD framework. Specifically, we propose a measurement-driven hybrid adaptive birth model for improved track initialization, employ adaptive detection probability parameters to effectively maintain tracks for occluded objects, and optimize density pruning and track extraction modules to further enhance overall tracking performance. Extensive evaluations on nuScenes and KITTI datasets show that OptiPMB achieves superior tracking accuracy compared with state-of-the-art methods, thereby establishing a new benchmark for model-based 3D MOT and offering valuable insights for future research on RFS-based trackers in autonomous driving.

💡 Research Summary

The paper tackles the problem of 3‑D multi‑object tracking (MOT) for autonomous driving from a model‑based perspective, proposing a novel random finite set (RFS) Bayesian filter called OptiPMB. While deep‑learning trackers have achieved impressive performance, they require large labeled datasets and lack interpretability. Traditional model‑based trackers rely on random‑vector (RV) filters (e.g., Kalman filters) within a tracking‑by‑detection (TBD) pipeline, but they depend on heuristic data‑association and complex track‑management logic. RFS‑based filters, such as the Poisson multi‑Bernoulli mixture (PMBM) and its approximation, the Poisson multi‑Bernoulli (PMB), provide a unified Bayesian framework that naturally handles object birth, survival, and death, yet existing PMBM/PMB implementations still suffer in dense, cluttered urban scenes because they use fixed birth intensities and static detection probabilities.

OptiPMB introduces three key innovations to overcome these limitations:

1. Measurement‑driven hybrid adaptive birth model – Instead of a static Poisson birth intensity λ_b, the method dynamically adjusts birth rates based on the current detection density and clustering of potential new objects. Candidate births are evaluated for spatial‑temporal consistency, and each receives a probability proportional to its likelihood of being a genuine new target. This reduces spurious tracks while quickly capturing genuine entrants, especially in highly dynamic intersections.

2. Adaptive detection probability (p_D) parameters – The filter estimates a per‑object visibility score using cues such as LiDAR point density, camera field‑of‑view coverage, and historical track confidence. The detection probability is then modulated: when visibility drops (e.g., due to occlusion), p_D is lowered to preserve the survival probability; when visibility recovers, the filter rapidly incorporates new measurements. This mechanism markedly reduces ID switches and premature track termination under occlusion.

3. Optimized density pruning and track extraction – During prediction, Bernoulli components with low existence probability r are pruned early, limiting memory and computational load. Remaining components are compressed into Gaussian mixture representations of high‑dimensional state (position, velocity, acceleration, size, orientation). In the extraction stage, a composite score combining existence probability and measurement likelihood selects the final set of tracks, ensuring robust performance without sacrificing real‑time capability.

The authors evaluate OptiPMB on the large‑scale nuScenes and KITTI benchmarks. On nuScenes, OptiPMB achieves an AMOTA of 0.767, surpassing previous model‑based methods (e.g., GNN‑PMB, RFS‑M3) and approaching the performance of state‑of‑the‑art learning‑based trackers. Improvements are especially pronounced in metrics related to ID maintenance (lower IDS) and handling of false positives/negatives in high‑clutter scenarios. Similar gains are reported on KITTI, where MOTA and ID‑switch rates improve significantly. Ablation studies confirm that both the adaptive birth model and the variable p_D contribute independently to the overall performance boost, and the system remains stable across a reasonable range of hyper‑parameters.

In summary, OptiPMB demonstrates that a carefully engineered RFS‑based filter can combine the interpretability, data‑efficiency, and theoretical soundness of model‑based approaches with tracking accuracy competitive with deep‑learning solutions. The paper suggests future work on multi‑sensor fusion (LiDAR, camera, radar), integration of learned association cues into the RFS framework, and deployment on embedded platforms for real‑time autonomous driving applications.