δ-Generalized Labeled Multi-Bernoulli Simultaneous Localization and Mapping with an Optimal Kernel-Based Particle...

Area of Science:

• Robotics and Artificial Intelligence
• Probabilistic Data Association and State Estimation

Background:

• Heuristic-based Simultaneous Localization And Mapping (SLAM) methods struggle with map and trajectory divergence under realistic conditions.
• Random Finite Set (RFS) frameworks offer solutions by circumventing external data association and managing sensor uncertainty.
• Existing RFS SLAM algorithms, like those using Probability Hypothesis Density (PHD) and Labeled Multi-Bernoulli (LMB) filters, suffer from information loss and particle degeneracy, leading to performance divergence.

Purpose of the Study:

• To develop an improved SLAM algorithm that addresses the limitations of current RFS-based approaches, specifically particle degeneracy and divergence.
• To enhance the accuracy and robustness of map and trajectory estimation in robotic SLAM.
• To leverage the computational advantages of efficient RFS filter variants within a particle filter framework.

Main Methods:

• Proposed a novel SLAM solution, δ-GLMB-SLAM2.0, integrating an optimal kernel-based particle filter with an efficient variant of the δ-Generalized Labeled Multi-Bernoulli (δ-GLMB) filter.
• Utilized the δ-GLMB filter to avoid information loss associated with LMB filter approximations in multi-target tracking and RFS SLAM.
• Evaluated the algorithm using simulated datasets and a portion of the KITTI dataset to assess performance against state-of-the-art methods.

Main Results:

• The proposed δ-GLMB-SLAM2.0 algorithm demonstrated superior performance compared to existing Rao-Blackwellized particle filter (RBPF)-based RFS SLAM algorithms.
• The algorithm achieved improved map quality and trajectory estimation accuracy, even when employing a limited number of particles.
• Results indicated a significant reduction in divergence from ground truth over time, addressing particle degeneracy issues.

Conclusions:

• δ-GLMB-SLAM2.0 offers a robust and accurate solution for robotic SLAM, particularly in challenging, realistic environmental conditions.
• The combination of an optimal kernel-based particle filter and an efficient δ-GLMB filter variant effectively mitigates common RFS SLAM limitations.
• The proposed method represents a significant advancement in RFS-based SLAM, outperforming current state-of-the-art algorithms.