An obstacle representation for off-road autonomous driving

We propose an object representation which allows the integration of range information with time, without the large computational and storage requirements of earlier representations. Potential obstacle regions in the image are described in terms of their range and direction with respect to the vehicle. Much effort has been put into matching Cartesian maps from frame to frame given the large uncertainty for distant objects. A directional representation however maintains the high accuracy of directional information (which is limited only by pixel density and camera calibration). Tracking objects from frame to frame is easier since the potentially large uncertainty in range does not corrupt the directional information. The use of this representation to identify untraversable regions is demonstrated on a series of real images taken from an off road vehicle