A Self-Supervised Learning Approach under Controller's Instruction for 3D Line Segment-Based Recognition of Semi-Unstructured Environment

It is difficult to achieve simple navigation in an unstructured environment, where the complexity of the environment and the amount of information the robot will have to process will make it impossible to navigate quickly. This paper presents a self-supervised approach to local path recognition for outdoor semi-unstructured environment, based on human instruction and automatic parameter selection method. Paths are be extracted to by means of 3D line segments by using the visual feature corresponding to 2D image and 3D point cloud. Visual odometry provided by RTAB-Map to determine the range of the robot's trajectory is used to obtain the line segments related to the local path. Human controller's instruction on the path for local navigation, together with the domain knowledge of paths, is reflected to parameter selection by Pareto optimality in the multiple objective optimization. It was verified that the proposed method could select an appropriate set of processing parameter tuples based on the knowledge and visual odometry. The proposed idea is expected to contribute selection of local paths for subsequent navigation in the robot's tracking with optimal parameter tuples.