Facade Reconstruction Using LiDAR Point Clouds And Panoramic Images Collected By Mobile Mapping Systems

With the development of sensors and rising demand for three-dimensional spatial information,3D city reconstruction has become a hot topic in research and application.Facade,as part of the building model,contains detailed and intuitive information.Mobile mapping system(MMS),as a multi-source sensor integrated mapping platform,is an effective way to collect spatial data on the resolution of street-level in urban areas.However,the current processing capability for mobile mapping data is not competent for the consistently growing massive amount of data acquired by mobile mapping systems.Therefore,there is still huge potential and sustainable value in the research of three-dimensional building modeling using mobile mapping data.Three-dimensional point cloud and image sequences are the two type of data collected by MMS.The characteristics of point clouds are as follows:firstly,they contains three-dimensional coordinates of the measured object surface;secondly,their spatial resolution is relatively low;thirdly,their measuring distance is limited.While the characteristics of the image sequences are:firstly,the contains rich,visual texture information;secondly their resolution are higher than point cloud,and accurate edge features can be extracted from them.This paper studies reconstruction of three-dimensional facade model using point clouds and images,which contains three key points,including the registration of point clouds and images,recognition of point cloud belonging to buildings,as well as three-dimensional modeling procedure of facades.The main contents of the whole thesis are as follows:(1)Based on the characteristics of the mobile point cloud and image sequences,we reviewed the related work at home and abroad,firstly covering the overall workflow and then focusing on the specific algorithms.Then we proposed our research goals,contents and organization of whole thesis.(2)We proposed a linear feature based registration method between mobile point cloud and panoramic image.The linear feature extraction algorithms for point clouds and panoramic images are discussed separately.Then the transformation model between point cloud and image was established based on linear constrains and ideal spherical model of panoramic camera,the solution process and analysis of accuracy are also given in detail.The matching procedures was designed for selecting corresponding lines and removing outliers.(3)We studied several typical segmentation algorithms for point cloud,and proposed two recognition methods.The basic concepts of neighborhood for the point cloud and the point feature were defined,and then we compared several segmentation algorithm for point cloud such as RANS AC,3D Hough transform and region growing.The two recognition algorithms are object-oriented and hierarchical respectively.The former one take segmentation results as objects,and calculated characteristic value of the particular class to decide whether the segments are belonging this class or not;the latter one is based on the observation that spatial distribution of point cloud are strongly correlated with its semantic information,and the hierarchical mechanism enables rapid recognition of point clouds belonging to defined classes.(4)We proposed coarse-to-fine facade modeling scheme using mobile point clouds and images.The main plane models of facades were firstly extracted from building point cloud.And we proposed BSP-tree based method to create depth model of facade:the outer boundary of point cloud segments are firstly regularized;then the region of facade are splitted by regularized line features into BSP blocks;finally the minimization of energy function method are applied to assign segments to BSP blocks,which guarantee the completeness of topological facade model.The windows detection methods based on point clouds and images are studied:firstly,we create the orthographic image by projecting panoramic image into main plane model of facades;then we detected window in point cloud using scanning line method,which provides alternative samples for the Joint Boost windows detection algorithms in images;and finally we merged the two results by rules.At last methods of texture mapping onto depth model of facades were studied.This paper studied the facade reconstruction using point clouds and images.For the key issues including registration between point cloud and image,recognition of point cloud,three-dimensional reconstruction of facades,the research has achieved good results.Further work will focus on fully automated corresponding line selecting,machine learning method for point cloud recognition,design of complex geometric facade model and 3D reconstruction using image sequence.With further research,the automation of building facades reconstruction using mobile LiDAR and image sequence will be improved.