Learning of dynamic environments by a mobile robot from stereo cues

Abstract

A system that builds a three-dimensional map of an indoor environment for a mobile robot is presented. The approach uses visual features extracted from stereo images as landmarks. A learning rule associated with each landmark is used to compute its existence state. New landmarks are merged into the map and transient landmarks are removed from the map over time. The location of the landmarks in the map is continuously refined from observations. The position of the robot is estimated by combining sensor readings, motion commands, and the current map state by means of an extended Kalman filter. The combination of neural network principles for map updating and Kalman filtering for position estimation allows for robust map learning of indoor dynamic environments.