Vision Based Far-Range Perception and Traversability Analysis using Predictive Probability of Terrain Classification

Abstract:
This paper describes a far-range visual map building and traversability analysis using geometrical information and image appearance, which are observed by a stereo camera, for autonomous mobile robots. In this research, long-range polar map that has multiple resolution of the radius direction is introduced, in order to use far area of terrain classification data effectively that has an huge error that grows quadratically with range. The traversability analyzing approach that reduces the influence of observation noise and a classification error is proposed. In this approach, appearance information is classified into traversable or not by Support Vector Machine to analyze from near to far area of traversability. Predictive probability of classification is calculated to reduce failure of terrain classification. Simultaneously, accurate analysis of traversability near a robot is accomplished by estimating planes of each grid from point cloud. Finaly, our system was experimented in outdoor environment.