Robot Localization in Dynamic Environments Using Time-Variant Semantic Cues

A challenging issue in the field of robotics is the lost robot problem, in which a robot has to relocalize itself in a previously mapped environment based on current sensor readings. We propose a method for addressing this problem by extending the mapping of an environment with semantically labelled points. These semantic landmarks are processed in an algorithm that registers two labelled point sets in order to obtain the rigid transformation that relates the robot's current frame with the global coordinate system. The relocalization system was tested on a dataset created for 3D scene analysis, and on self-made scans of several environments with different types of visual interferences, including obstructions to the field of view and severe lighting changes. These tests revealed that the proposed method performs well with realistic levels of noise in the input data. It even outperforms a state-of-the-art visual mapping and relocalization system both in robustness and in accuracy, producing linear and angular errors of less than 10 cm and 1 degree respectively in successful relocalization attempts