3D Building Model Reconstruction Based On Improved Image Segmentation And MS-SIFT Algorithm

The reconstruction of the 3D model of the building is of great significance in the fields of urban planning,virtual tourism,and protection of ancient buildings.The image processing algorithm is low in cost,easy to operate,time-saving and labor-saving,suitable for 3D reconstruction of large buildings.However,most of the buildings are in a more complex environment.When using image processing algorithms for reconstruction,they will be interfered by many external factors,and the 3D reconstruction algorithm itself is not mature.Therefore,the 3D reconstruction of buildings has become a hot and difficult problem in current research.In common images containing buildings,in addition to buildings,there are many other elements,such as trees,grass,etc.These interference items that are not related to buildings often have extreme points in many pixels,which will cause great interference to the accuracy of feature point detection and matching.Of course,there are also many weak texture areas on the wall of the building.It is not easy to detect feature points in these weak texture areas.As a result,only a small amount of point clouds can be obtained after reconstruction,resulting in a large area of holes missing,which affects the quality of the final reconstruction model.Therefore,removing interference information,dealing with the detection difficulties of weak texture areas,and using better reconstruction methods will be the key to improving the quality of the building reconstruction model.The research content of this thesis is the research and implementation of 3D model reconstruction methods using multiple building images from different perspectives.First,perform feature point detection and matching on the input images from different perspectives to estimate the perspective conversion relationship of different images;Secondly,use the multiview geometric relationship and the triangulation algorithm to map two-dimensional points into the space to obtain the three-dimensional space of the feature points coordinates;Finally,a dense point cloud of the building is obtained through an incremental reconstruction method.The main work content of this thesis includes the following aspects:(1)Propose a building image segmentation algorithm combining visual saliency and connected domain segmentation.In the process of building 3D reconstruction,many feature points will be detected on nonbuilding targets such as bushes,trees and background,which will interfere with the subsequent matching process and affect the matching time and accuracy.In response to these problems,this thesis propose a image segmentation algorithm for building subjects combining visual saliency and connected domain segmentation.This thesis propose a building image segmentation algorithm.Firstly,the visual saliency algorithm and the connected domain segmentation algorithm are used to obtain the rough building and complete the preliminary segmentation.Then,the filtered background and the remaining building body are divided into blocks,and the color vector,texture vector and a building super green factor suitable for the building scene are used to remove the remaining background block in the preliminary segmentation result.Finally,use median filtering to solve the problem of rough edges of the segmented image,and then use a connected domain segmentation algorithm to keep the main part of the building intact and remove useless pixels to complete the final segmentation.This thesis compares with two better segmentation algorithms from subjective visual observation and objective evaluation indicators.Experiments show that the algorithm in this thesis does not require complicated human-computer interaction,is not afraid of complex scenes,and is not sensitive to light.It can obtain the final segmentation results as accurately as possible under complex scenes and strong light.(2)Propose a feature point extraction algorithm that is also applicable to weak texture regions.Many existing feature point detection algorithms mostly improve the performance of the algorithm running time and maintain the accuracy of feature points,but there are relatively few studies on building scenes.In response to the above problems,this thesis compares many existing feature point detection algorithms from subjective visual observation and objective data,and propose a feature point detection algorithm suitable for 3D reconstruction of buildings.Aiming at the limitations of existing feature point detection algorithms in weak texture regions,this thesis propose a suitable for multiple scenarios scale-invariant feature transform(MS-SIFT)algorithm.This thesis compares the detection effects of building feature points from the subjective visual observation and objective data of various existing feature point detection algorithms,and propose a feature point detection algorithm which is suitable for 3D reconstruction of buildings.At the same time,in view of the limitations of the existing feature point detection algorithms in weak texture areas,a MS-SIFT algorithm is proposed,which improves the detection performance of weak texture area in three aspects: enhanced texture,feature point detection strategy adjustment and introduction of new decision conditions,and compared with the original scale-invariant feature transform(SIFT)algorithm in terms of subjective visual observation and objective evaluation indicators.Experiments show that the MS-SIFT algorithm in this thesis can improve the limitations of the existing feature point detection algorithm in weak texture areas,the number of effective points in both texture rich areas and weak texture areas is increased,the detection effect of weak texture areas is better,the total number of feature points is increased and algorithm performance is improved.(3)Propose a method of building 3D reconstruction based on improved image segmentation and MS-SIFT algorithm.Aiming at the problems that the existing 3D reconstruction methods have less research on building and the reconstruction effect is not ideal caused by the immature 3D reconstruction algorithm,this thesis combines the above two methods to provide a 3D reconstruction method of buildings based on improved image segmentation and MS-SIFT algorithm to achieve 3D reconstruction of buildings.This thesis uses the given building 3D reconstruction process,adopts the incremental reconstruction method,introduces the idea of image segmentation,and obtains an image that only contains buildings.Then in the structure-from-motion(SFM)method,the MS-SIFT algorithm of this thesis is used to obtain more feature points containing weak texture area points,and these feature points are used to obtain the sparse point cloud of the building,and then the point cloud is densified to obtain the denseness of the building Point cloud,realize the whole process.After the reconstruction is completed,this thesis compares with the original 3D reconstruction algorithm from subjective visual observation and objective evaluation indicators.Experiments show that the reconstruction method in this thesis can reduce the useless points on the basis of ensuring the accuracy of the original algorithm,keep only the key information of the building as much as possible,and increase the number of effective feature points,improve the large-area hole missing phenomenon in the 3D reconstruction of the building,the number of dense point clouds of the obtained model increases significantly,greatly improves the accuracy of the building reconstruction model,and can obtain more accurate dense point clouds and 3D models of buildings.Experimental results show that the method proposed in this thesis can effectively reduce the impact of non-target objects,improve the phenomenon of missing holes after the 3D reconstruction of the building,and enhance the effect of the reconstructed 3D model of the building.