GPU-accelerated Localization in Confined Spaces using Deep Geometric Features

Abstract

Navigating within dark and confined spaces require robotic platforms to utilize accurate and reliable localization systems to operate safely and unattended. This paper presents an absolute localization system, for known confined spaces, using state of the art 3D pointcloud descriptors. Local geometric features are extracted from a known map and registered to matching features visible in the robots field of view. The 3D registrations are motion-filtered and fused with a visual inertial odometry estimate in an Extended Kalman filter, which return a fast and accurate absolute pose estimate. The proposed localization system is tested with different deep learning feature descriptors in a structured confined space, and our results indicate greater accuracy and lower processing time when compared to mainstream 3D registration approaches.