3D Point Cloud Reconstruction Of Underground Space Based On Monocular DSO

With the progress of urbanization in China,the land sources on the ground are becoming increasingly scarce.Development of the city has been vigorously expanded underground,and the utilization rate of underground space is getting higher and higher.However,because of the characteristics of the underground environment such as hermeticity,load-bearingness and invisibility,safety inspection is particularly important.Traditional methods of manual inspection require a great deal of manpower and time cost,and it is difficult to meet the growing underground space development needs of the city.In recent years,inspection robot technology represented by machine vision has developed rapidly,whitch has greatly relieved the task of manual inspection.But at present,the inspection route of the inspection robot is mainly set by human beings and cannot be applied to uncertain environments.Afterwards,existing autonomous inspection and navigation methods,represented by neural networks,still require a large amount of time and samples to train,so this kind of method has not wide applicability.Therefore,combining achievements of SLAM(Simultaneous Localization and Mapping)developments,an improved 3-D reconstruction method based on monocular DSO(Direct Sparse Odometry)is proposed to deal with this problem.Works of this thesis are mainly reflected in the following three aspects.Firstly,building data acquisition platform to acquire sample data.Then,an image enhancement method based on RETINEX theory is proposed to solve the problem of image quality inconsistency caused by the inhomogeneity of illumination in the underground environment.Validity of proposed method is verified by examples to ensure the consistency of image quality,which provides a good data source for the implementation of subsequent.Secondly,in order to solve the problem of point cloud size uncertainty in monocular DSO 3-D reconstruction,a point cloud scale determination method based on coordinate transformation is proposed.Experiments show that proposed method can be effectively applied to the size determination in monocular 3-D reconstruction.Finally,using the c/c++,on the Ubuntu platform,a 3-D point cloud reconstruction prototype system for underground space based on monocular DSO is designed and implemented.The system integrates proposed methods in this thesis and can systematically realize functions from data acquisition to 3-D reconstruction.By comparing with the original DSO method,the proposed method can clearly get the target 3-D reconstruction,and the point cloud data is less.At the same time,the required hardware is only a monocular fixed-focus camera,which is less costly and solves the problem of scale uncertainty of monocular 3-D reconstruction.