Absolute localisation in confined spaces using deep geometric features

Abstract

When operating in dark and confined spaces, the capacity of the robots to localise in an absolute reference frame is of utmost importance. This letter presents an absolute localisation system, using deep 3D landmarks, for known confined spaces. The system estimates the robot's relative localisation using visual inertial odometry. Local deep 3D landmarks are extracted from the robot's view. Similar 3D landmarks are, also, extracted from the prior map which are then registered with the local landmarks to provide absolute localisation via an extended Kalman filter. To the best of knowledge, deep 3D feature registration has not been used before for absolute localisation. The proposed localisation system is tested within a representative application area—i.e. a structured, confined space—and the results indicate greater accuracy and lower processing time when compared to mainstream 3D registration approaches.