Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MegaParticles: Range-based 6-DoF Monte Carlo Localization with GPU-Accelerated Stein Particle Filter

Kenji Koide, Shuji Oishi, Masashi Yokozuka, Atsuhiko Banno

TL;DR

The paper tackles robust 6-DoF sensor pose localization from 3D range data, addressing global re-localization without an initial pose and resilience to sensor dropouts. It introduces a GPU-accelerated Stein variational gradient descent (SVGD) framework with a Gauss-Newton approximation, a locality-sensitive hashing (LSH) based neighbor search in $SE(3)$, and a dynamic neighbor-graph Bayesian filter for posterior propagation, enabling real-time updates of up to $10^6$ particles on a single GPU. Key contributions include an efficient nearest-neighbor field for likelihood evaluation, scalable $SE(3)$ neighbor search, and a posterior propagation scheme that handles multi-modal distributions without resampling. Experimental results in indoor and outdoor settings demonstrate strong kidnapping robustness and rapid re-localization, highlighting the method’s scalability and practicality for autonomous systems in challenging environments.

Abstract

This paper presents a 6-DoF range-based Monte Carlo localization method with a GPU-accelerated Stein particle filter. To update a massive amount of particles, we propose a Gauss-Newton-based Stein variational gradient descent (SVGD) with iterative neighbor particle search. This method uses SVGD to collectively update particle states with gradient and neighborhood information, which provides efficient particle sampling. For an efficient neighbor particle search, it uses locality sensitive hashing and iteratively updates the neighbor list of each particle over time. The neighbor list is then used to propagate the posterior probabilities of particles over the neighbor particle graph. The proposed method is capable of evaluating one million particles in real-time on a single GPU and enables robust pose initialization and re-localization without an initial pose estimate. In experiments, the proposed method showed an extreme robustness to complete sensor occlusion (i.e., kidnapping), and enabled pinpoint sensor localization without any prior information.

MegaParticles: Range-based 6-DoF Monte Carlo Localization with GPU-Accelerated Stein Particle Filter

TL;DR

The paper tackles robust 6-DoF sensor pose localization from 3D range data, addressing global re-localization without an initial pose and resilience to sensor dropouts. It introduces a GPU-accelerated Stein variational gradient descent (SVGD) framework with a Gauss-Newton approximation, a locality-sensitive hashing (LSH) based neighbor search in , and a dynamic neighbor-graph Bayesian filter for posterior propagation, enabling real-time updates of up to particles on a single GPU. Key contributions include an efficient nearest-neighbor field for likelihood evaluation, scalable neighbor search, and a posterior propagation scheme that handles multi-modal distributions without resampling. Experimental results in indoor and outdoor settings demonstrate strong kidnapping robustness and rapid re-localization, highlighting the method’s scalability and practicality for autonomous systems in challenging environments.

Abstract

This paper presents a 6-DoF range-based Monte Carlo localization method with a GPU-accelerated Stein particle filter. To update a massive amount of particles, we propose a Gauss-Newton-based Stein variational gradient descent (SVGD) with iterative neighbor particle search. This method uses SVGD to collectively update particle states with gradient and neighborhood information, which provides efficient particle sampling. For an efficient neighbor particle search, it uses locality sensitive hashing and iteratively updates the neighbor list of each particle over time. The neighbor list is then used to propagate the posterior probabilities of particles over the neighbor particle graph. The proposed method is capable of evaluating one million particles in real-time on a single GPU and enables robust pose initialization and re-localization without an initial pose estimate. In experiments, the proposed method showed an extreme robustness to complete sensor occlusion (i.e., kidnapping), and enabled pinpoint sensor localization without any prior information.
Paper Structure (14 sections, 7 equations, 22 figures, 2 tables, 1 algorithm)

This paper contains 14 sections, 7 equations, 22 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Iterative Scan Matching-based Localization
  4. Monte Carlo Localization
  5. GPU-Accelerated Localization
  6. Methodology
  7. Problem Setting
  8. Prediction Step
  9. Correction Step
  10. Posterior Propagation on Neighbor Particle Graph
  11. Experiment
  12. Indoor Experiment
  13. Outdoor Experiment
  14. Conclusion

Figures (22)

  • Figure 1: (a) Proposed method performs 6-DoF sensor localization with one million particles. All the particles are evaluated and updated in real-time on a single GPU. Point clouds acquired by a MS Azure Kinect are used (No IMU input). (b) Posterior probability distribution. (c) A close look at the maximum posterior particle (zoom in to see RGB-colored pose particles), and (d) the same view with particles colored based on the posterior probabilities.
  • Figure 2: SE3 locality sensitive hashing based on a stable distribution.
  • Figure 3:
  • Figure 4:
  • Figure 5:
  • ...and 17 more figures