Computer vision algorithms for autonomous mobile robot map building and path planning

Abstract

Describes an algorithm for autonomous mobile robot map building and path planning in an unknown environment. One potential application is that of the AMR transporting an individual in response to a verbal command. i.e. an intelligent wheelchair. The sensor system used by the robot is stereo vision. We use a view-based, probabilistic approach, whereby the AMR learns its environment as a series of visually different locations. The algorithm is particularly robust, and notably the robustness does not vary with the size of the learnt environment. This enables large environments to be learnt reliably. This was previously an unsolved problem. The algorithm enables the AMR to learn an unknown environment without the need for internal distance measurements (odometry), using feature observations only. The algorithm allows an AMR to learn a new environment with a single passing and once the environment has been mapped, its topology can be inferred from the stored map. Hence all possible paths through that environment are navigable, even paths previously untraversed. Furthermore, using a limited amount of real-world distance information, derived from stereo cameras, the robot is able to construct a robust internal representation of its position with respect to visible features, and thus perform accurate intra-region localisation.