Outdoor Global Localization via Robust Registration of 3D Open-Set Segments

Mason Peterson1, Yi Xuan Jia1, Yulun Tian1,2, Annika Thomas1, Jonathan How1,
1Massachusetts Institute of Technology 2University of California, San Diego
Paper arXiv Code Data
Example Image

Pair of segment submaps matched by two robots traveling in opposite directions in an off-road environment. The submaps are shown in different colors (red and blue) together with robot trajectories. Green edges denote segment-level associations found by the proposed method, and a subset of these associated segments are annotated in the corresponding image observations.

-->

Abstract

Global localization is a fundamental capability enabling long-term and drift-free robot navigation. This paper presents an outdoor global localization method based on robust registration of open-set segment maps. Our method detects and tracks segments using open-set image segmentation models that enable direct generalization to unseen environments. To perform global localization under the high outlier regimes that are typical of natural/outdoor environments, we formulate a registration problem between small submaps of 3D segments and solve for the correspondences using a graph-theoretic global data association approach. Further, to guide registration in highly noisy or ambiguous scenarios, we propose novel ways of incorporating additional information (e.g., segment attributes and known gravity direction) within the global data association formulation. The proposed method is evaluated on outdoor datasets recorded by multiple robots and shown to outperform existing methods in terms of precision-recall metrics and localization accuracy.

-->

BibTeX

@article{peterson2024roman,
  author    = {Peterson, Mason and Jia, Yi Xuan and Tian, Yulun and Thomas, Annika and How, Jonathan},
  title     = {Outdoor Global Localization via Robust Registration of 3D Open-Set Segments},
  journal   = {arXiv preprint [add here]},
  year      = {2024},
}