Stereo Vision-Based Navigation in Unknown Indoor Environment

Abstract

Different applications in the field of vision-based navigation of autonomous mobile robots depend on the degree of knowledge of the environment. Indoor environment applications often use landmarks or maps for navigation. Others have only knowledge of known and expected objects. In such applications, parts of the scene are classified in these objects, e.g. road junctions, doors, walls, furniture, and a possible path will be estimated. In case of a lack of a priori knowledge of the environment, we propose an approach for vision-based navigation, considering any reconstructed 3D point of the scene. A line segment stereo algorithm and a reconstruction procedure lead to uncertain 3D points of the scene in front of the mobile system. All these 3D points are regarded as obstacles and a following trace estimation will be applied on this 3D data. In order to increase the reliability of reconstructed 3D points, a validation step excludes impossible 3D points, exploiting the stereo geometry of the vision system. After validation a two-step analysis is applied, which contains the minimum distance method and point distribution analysis method. This analysis leads to a possible trace for the mobile robot, resulting in values for steering angle and velocity given to the mobile system. The method has been implemented on the experimental system MOVILAR (MObile VIsion and LAser based Robot) which is based on a multi-processor network. It achieves actually process cycles of approximately 1.5s and a velocity of about 10 cm/s, i.e. a slow walking speed. Experimental results of the above mentioned methods are presented.