Line Enhanced Monocular 3D Reconstruction For Urban Environments

The prerequisite for a mobile robot to perform tasks autonomously in an unknown environment is to locate itself and build an environmental map based on the information acquired by on-board sensors.The reconstructed map can be used for the following navigation,localization,path planning,and visualization.GPS sensors limit its applications due to accuracy,and secondly,it is difficult to receive GPS signals in indoor environments.Camera is one of the most commonly used sensors,which can quickly obtain the visual information of the scene and is widely used in computer vision and robotics.This paper tackles the problem of autonomous behavior requirements for mobile robots using a purely monocular visual based method,and mainly researches on two aspects: autonomous localization of mobile robots and reconstruction of environmental maps.The main contributions of this thesis are as follows.(1)Aiming at the serious degeneracy in three-dimensional line reconstruction,a new minimum spatial line representation is proposed for the uncertainty analysis of line reconstruction.The proposed uncertainty analysis summarizes the main factors affecting the line reconstruction and is capable of providing quantitative evaluation of reconstruction accuracy.In addition,a three-dimensional line map reconstruction method suitable for Sf M and SLAM post-processing is proposed,and the line map is culled by the proposed uncertainty analysis.Experimental results show that the reconstructed three-dimensional line map outperforms sparse point cloud in sense of structure.(2)In consideration of the fact that urban scenes usually contain a large number of texture-less areas and fewer feature points but present abundant line segments,a real-time monocular SLAM that integrates both point and line features are proposed.A higher level of line features is extracted to constrain pose estimation in texture-less scenes and a3 D line map is maintained online,the SLAM accuracy is improved by line features constraints.Specifically,a multi-step Sf SM for small baseline with points and line segments is proposed to obtain a more accurate inital map for SLAM system.(3)The surface model generated by sparse point cloud inevitably suffer from structural distortion,especially at the edges.To this concern,a surface model generation method enhanced by line segment is proposed to maintain the structure of the scene at the edges.For the online local surface model generation,a two-dimensional constrained Delaunay Triangulation on the point and line features is performed on the image,then the mesh is back projected into three-dimensional for optimization and smooth.For the offline global surface model generation,the sparse point cloud is densified to enhance the details with prior pose estimation and the line cloud is rectified by the ”Manhattan World” assumption to constrain three-dimensional Delaunay Triangulation,the surface model generated by the advancing front method is not only rich in details but also good in structure sense.(4)Aiming at the problem of three-dimensional reconstruction of urban scenes,an urban scene reconstruction system based on low-altitude UAV images is designed.The reconstruction system can be divided into three modules: data capturing,data transmission and data processing.The UAV is controlled by manual operation or preset route to capture images of target scene,and the captured images are transmitted to the ground station by microwave in real time.The system builds a sparse map consists of point and line landmarks on the fly and an edge-preserving global surface model offline.