Indoor Global Localization Using Depth-Guided Photometric Edge Descriptor for Mobile Robot Navigation

This paper suggests a new landmark descriptor for indoor mobile robot navigation with sensor fusion and a global localization method using it. In previous research on robot pose estimation, various landmarks such as geometric features, visual local-invariant features, or objects are utilized. However, in real-world situations, there is a possibility that distinctive landmarks are insufficient or there are many similar landmarks repeated in indoor environment, which makes accurate pose estimation difficult. In this work, we suggest a new landmark descriptor, called depth-guided photometric edge descriptor (DPED), which is composed of superpixels and approximated 3D depth information of photometric vertical edge. With this descriptor, we propose a global localization method based on coarse-to-fine strategy. In the coarse step, candidate nodes are found by place recognition using our pairwise constraint-based spectral matching technique, and the robot pose is estimated with a probabilistic scan matching in the fine step. The experimental results show that our method successfully estimates the robot pose in the real-world tests even when there is a lack of distinctive features and objects.