Cooperation Between Stereo-vision And A Token Tracker Process For The 3d Vision Of A Mobile Robot

Abstract

We present a 3D vision sensor for a mobile robot. The sensor builds a description of the environment based on 3d vertical straight line segments. It operates as a cooperation between a stereovision process and a token tracking process. A full matching process is performed at the initialization step. Then the token tracker allows to maintain the same matches that were established previously. The sensor has only to find out matches on the new tokens lhat were detected. We use a procedure based on the dynamic programming to find out these new matches. This procedure takes advantage of the already existing matches to perform this task rapidly. 1- IntrodA mobile robot has to construct a description of its environment €or the navigation task. This description has to be built continuously and should not enforce the robot to stop during this task. We present a 3D stereo-based sensor that will be on board of a mobile robot used for surveillance of industrial sites. It constructs a description of the environment based on 3d vertical straight line segments. The restriction to vertical straight lines is reasonable in the sense that the sensor will be used only for the navigation task. We aim to reduce the complexity and to increase the speed of the algorithms. Indeed (i) the amount of segments to be processed is reduced, (ii) the verticality hypothesis drastically simplifies the algorithms. Anyway, the poor description of the environment will be fused with depth informations provided by the sonar sensors. -171 The functioning of the sensor is based on the combination of two processes : a stereo matching process and a token tracking process. The stereo matching process is useful for the 3D reconstruction. It embodies the actual 3D capability of the sensor. The token tracking process deals with the image changes due to the robot motion. The main problem for the stereo matching process comes from the socalled stereo - c or r e s p o n de nc e problem. Robustness is often paid by a very time consuming matching process. The basic idea of the sensor described in this paper is to take advantage of the token tracking process to aid the stereo process. It relies on the previous detected correspondences to rapidly compute those of the current image stereo pair.