Research And System Implementation Of Monocular Obstacle Avoidance In Indoor Scene

In the field of robotics and autonomous driving,obstacle detection and avoida nce are basic safety technologies,which helps unmanned vehicles to perceive objects in the scene and timely controls the vehicle to avoid obstacles.In recent years,with the increasing research of mobile robots and autonomous driving,this technology has also received extensive attention from academia and industry.Inspired by the perception mechanism of human using vision,this paper proposes an obstacle detection method based on a multi-layer framework.The method combines the three-dimensional spatial cue of the scene(i.e.,ground homography)with the two-dimensional appearance cue of the obstacle(i.e.,occlusion edge,color,etc.),so that it can detect obstacles in a variety of complex scenes.The research in this paper is divided into four parts:obstacle-aware occlusion edge extraction,spatial visual occlusion edge extraction,obstacle-aware regression,the design and implementation of obstacle avoidance system.(1)Obstacle-aware occlusion edge extraction.This thesis establishes a multi-layer framework and expounds the extraction process from three aspects:multi-layers regions generation,multi-distance edge cues fusion and occlusion edge generation.Among them,the focus is to extract the multi-layers regions related to obstacle distance,and design the method of edge cues fusion to enhance the edge of the obstacle,so that the occlusion edge can clearly fit the contour of all obstacles.(2)Spatial visual occlusion edge extraction.In this thesis,spatial visual occlusion edges are the edges above the ground,which are extracted by using the obstacle contours observed from multiple position.The method is elaborated from three main aspects:ground homography estimation,homography error and spatial edge points extraction.The homography matrix estimation and the homography error calculation are the key points.Finally,qualitative experiments are performed on the indoor scene dataset.(3)Obstacle-aware regression.This thesis describes the regression model from three aspects:sample extraction,feature selection and regression construction.The focus of this section is on the design of pre-processing of samples using pseudo-distance and the hand-crafted features suitable for obstacle detection.This ensures that obstacles are discovered as much as possible.Finally,experimental evaluations are performed on the indoor scene dataset and the outdoor Lost and Found dataset.(4)Design and implementation of obstacle avoidance system.In this thesis,the above algorithms are used to detect the obstacles,and an obstacle avoidance system is designed to make the robot autonomously avoiding obstacles.