セマンティック・セグメンテーションによるシニアカーの走行可能領域検出手法の開発

抄録

In the development of an automatic driving system of mobility scooters, a traversable area detection method in the range of about 10m where position accuracy can not be guaranteed by GPS is important. We propose a traversable area detection method that can perform advanced traversable area recognition using semantic segmentation. Semantic segmentation classifies images into several meaningful classes on a pixel by pixel basis. Therefore, we make data set classified objects on the forward image into three classes based on the judgment criteria of traveling recommendation, and constructed a system that can classify objects online using SegNet. By matching a classified image with point cloud of a stereo camera and updating a cell of the traversable recommendation degree map by bayesian filter, it is calculated how much traveling is recommended for mobility scooters on road surface. We tested the system in a road scene. As a result, it was found that plausible traversable area can be obtained by the proposed method.