ミーンシフト法を用いた特徴選択に基づく危険源検出

抄録

A method to detect hazards in the sense of traversability on pedestrian paths is presented for the purpose of supporting hazard recognition by robotic transporter users. Target transporters are supposed to be equipped with a 3D range image sensor to observe potential hazards like road roughness, drivability, and positive/negative obstacles. Hazard detection problem is formulated as a supervised learning such that point cloud data sets labeled hazard or non-hazard by a transporter user are input to the learning system and the learned output produces a hazard prediction to a novel data set. The present study tackles a reliable feature selection from a clustered point cloud with observation density inversely proportional to the square of the distance and found that feature selection by mean shift clustering works better than those by graph based clustering. Experiment results show that the proposed method is able to discriminate hazard cases from non-hazard ones of traffic control posts, walls, and down steps along pedestrian paths that previous approaches fail to predict correctly. The processing time of those predictions meets the requirements for online hazard detection.