Combination of Geometrical and Statistical Methods for Visual Navigation of Autonomous Robots

摘要

For visual navigation of an autonomous robot, detection of collision-free direction from an image/ image sequence captured by imaging systems mounted on the robot is a fundamental task. This collision free direction provides the next view to direct attention for computing the next collision free direction. Therefore, the robot requires a cyclic mechanism directing attention to the view and computing the collision free direction from that view. We combine a geometric method for free space detection and a statistical method for visual navigation of the mobile robot. Firstly, we deal with a random-sampling-based method for the detection of free space. Secondly, we deal with a statistical method for the computation of the collision avoiding direction. The robot finds free space using the visual potential defined from a series of views captured by a monocular camera system mounted on the robot to observe the view in front of the robot, We examine the statistical property of the gradient field of the visual potential. We show that the principal component of the gradient of the visual potential field yields the attention direction of the mobile robot for collision avoidance. Some experimental results of navigating the mobile robot in synthetic and real environments are presented.