A real-time obstacle detection vision system for autonomous high speed robots

This paper describes a real-time obstacle detection system for automobile-class robots using 3D point clouds generated from a commercial stereo camera pair. This approach is geared towards autonomous vehicle navigation for unstructured prescribed outdoor environments (e.g., off-road) as defined by the DARPA Grand Challenge, but will also work on structured environments (i.e., roads). The obstacle detection algorithms use various slope and elevation calculations to determine the presence of a non-passable area. Unlike many previous systems that assume that the path is a flat surface, this system also performs well on non-flat surfaces, which is a requirement for off-road conditions. Calculations are done in relative space using a grid system to map out traversable and nontraversable grid coordinates. Algorithms are developed to filter out erroneous points and extract features from the 3D point cloud for object classification for areas of limited texture. The overall system has been tested at speeds up to 30mph (14m/s) on the University of Louisiana's CajunBot II (Jeep Wrangler) robot.