人手による画像上への経路指示を利用したスコアの学習に基づく妥当な走行領域の推定

抄録

This paper describes a novel method of traversable region estimation in a variety of scenes. Our estimation targets are traversable areas where human selects in common sense: e.g. the side of a sidewalk and an animal trail. In the proposed method, a scene image is first segmented into categories such as roads and grass, and then traversable regions are estimated by considering the layout of the categories. In the region estimation, human common sense modeled using learning data is used. In learning phase, the operator instructs possible paths, for instance, “moving on the sidewalk along guardrails”, “crossing a crosswalk”, and “avoiding obstacles”. Specifically, the operator draws polygonal lines as paths on each image. In this phase, one important thing for the operator is to consider traffic rules, object avoidance, accident prevention and so on. Since traversable regions are estimated based on a learning result including operators' common sense, they contain generalized common sense which is suitable for mobile robots moving in everyday environments. Experiment shows that this method is successful in generating the correct estimation in a variety of scenes.