Traversable Region Estimation for Mobile Robots in an Outdoor Image

This paper describes a novel method to estimate appropriate traversable regions from an outdoor scene image. The traversable regions output by the proposed method reflect the common sense of people. For example, a candidate traversable region is "a paved road somewhat distant from the side ditch." The input to the traversable region estimation is one color image. First, category is assigned to each pixel in the image. The categorization result is then input to the region estimator. Finally, the traversable region are estimated on the input image. An important aspect of this method is the application of two score functions in region estimation process. One score function places high value on categories selected as traversable paths by subjects. The other function places high value on categories that are not selected as traversable regions but are adjacent to categories with traversable paths. A combination of these two functions produces feasible estimation results. The effectiveness of the combined score functions was evaluated by experiments and a questionnaire.