Integrated subsystem for Obstacle detection from a belt of micro-cameras

This paper describes the on-going work on the design and the implementation of an integrated visual system dedicated to the classical function Obstacle detection. This system will be embedded on a mobile robot, which must navigate in a cluttered and dynamic environment: typically a public area like pedestrian streets, a transport station or a commercial center. The current robots are equipped typically by a belt of ultrasonic sensors or by Laser Range Finders: it is proposed here to evaluate how a set of micro-cameras mounted around a mobile robot could be used in order to detect free space and obstacles. During an off line learning step, the appearance-based characteristics of the ground are extracted and recorded; then on line, images are acquired synchronously by micro-cameras, every pixel on every image is classified as Ground or Obstacle from color and texture attributes, and geometrical constraints between successive images are applied in order to validate that detected obstacles are above the ground. Finally all information are fused in a single robot-centered occupancy grid. This paper presents algorithms proposed for ground classification and obstacle validation, and describes first results about the architecture proposed for the integrated system.