Visual Detection of Obstacles Assuming aLocally Planar

Obstacle avoidance is an essential capability of an autonomous robot. This paper presents a method that enables a mobile robot to locate obstacles in its eld of view using two images of its surroundings. The method is based on the assumption that the robot is moving on a locally planar ground. Using a set of point features (corners) that have been matched between the two views using normalized cross-correlation, a robust estimate of the homography of the ground is computed. Knowledge of this homography permits us to compensate for the motion of the ground and to detect obstacles as areas in the image that appear nonsta-tionary after the motion compensation. The resulting method does not require camera calibration, is applicable either to stereo pairs or to motion sequence images, does not rely on a dense disparity//ow eld and circumvents the 3D reconstruction problem. Experimental results from the application of the method on real images indicate that it is both eeective and robust.