LiDAR-Interacted and Self-Guided Further Scene Parsing for Road Detection

Road detection is a crucial task for unmanned systems to assess the drivable area and avoid the obstacles and potholes to ensure vehicle safety. Despite the sparse data, LiDAR could provide accurate height and distance clues in all road conditions, which is an incremental complement to enhance the detection limited by camera sensors with rich visual features. In this paper, transformed 2D LiDAR data is used as a confident mask and interacts with visual images to cut down useless visual semantic information captured by cameras. After acquiring detection result from scene parsing network with interacted visual data and LiDAR data, a second parsing stage is adopted which further pushes the proposed model to preserve better information and thus gives more robust road identification. We report the experimental results on KITTI dataset to show its effectiveness and efficiency.