Free Space Detection from Catadioptric Omnidirectional Images for Visual Navigation using Optical Flow

Institute of Media and Information Technology, Chiba University, JapanYayoicho 1-33, Inage-ku, Chiba, 263-8522, JapanAbstract. In this paper, we develop a free space detection algorithm forthe visual navigation of the autonomous robot mounting a catadioptiricomnidirectional imaging system. The algorithm detects the dominantplane as the free space from a sequence of omnidirectional images cap-tured by a camera mounted on the autonomous robot. The dominantplane, which can be detected from the optical-°ow ¯eld, is the largestplanar area in the image. For the detection of the dominant plane fromthe optical-°ow ¯eld, we adopt the motion separation property, that is,the optical-°ow vector is decomposed into in¯nitesimal rotation, trans-lation, and divergent motions on the images. The algorithm matchesthe measured translation optical-°ow ¯eld with the template translationoptical-°ow ¯eld to separate the dominant-plane as the free space forthe navigation and the obstacle area. The template optical-°ow ¯eld isgenerated from a preobserved image sequence without any calibration ofthe internal parameters of both the robot and camera.