Structure Deformation Monitoring Based On Point Cloud Scene Feature Registration

Deformation refers to the change of shape,size and position of an object under the influence of external factors.There are a large number of uncertain factors affecting the structural safety of structures,such as tidal phenomena,groundwater changes,hurricane disasters,and geological disasters,monitoring the deformation of such structures is particularly important to ensure the safety of people's lives and property.The terrestrial 3D laser scanning technology can reflect the overall deformation trend of the structure and can achieve comprehensive disaster warning.The TLS technology is used to monitor the deformation,mainly involving data preprocessing,data registration and expression of deformation variables.Common point cloud registration methods include the use of targets,the use of point cloud image features,and direct registration using the point cloud itself,the target is easily destroyed,and the image characteristics of the point cloud image require additional image or reflection intensity information,the direct registration is susceptible to the change region,and each method has its own limitations.And there is no uniform expression standard for the deformation solution.In this paper,the point cloud scene feature is used to automatically identify the stable region for registration.Based on the nearest point and mesh model,the distance between the point and the nearest point local normal vector is used as the expression of the shape variable.The work is as follows:(1)For the redundancy of point cloud data,streamline point cloud data with the bounding box method,for outlier noise points,combine statistical outlier removal and radius outlier removal to eliminate such outliers;(2)For the point cloud data,the voxel is used as the base element,and the feature vector set of the voxel is constructed by using the geometric features and the spin image features,the single point feature vector is clustered by K-means to construct a word dictionary,and the word bag model is used to normalize the voxel feature vector set,using the 4-points congruent sets algorithm to match the voxels of different periods of data through the voxel center of gravity.Using cosine similarity to judge the similarity of conjugate voxel eigenvectors,eliminate low similar regions,and use high similar regions,ie stable regions,to complete registration;(3)According to the roughness of the reference data point cloud,the appropriate neighborhood is determined to solve the local normal of the point cloud and adjust the normal direction.Using the k-d tree to search for neighbor points and solving the projection distance of the Euclidean distance of the adjacent point in the normal direction,the shape variable can be obtained.The experiment in this paper contains simulation data and real data,each data contains two phases: stable data and change data.The experimental results show that: Using the scene features to automatically identify the stable area,the simulation data can avoid the manually moved bricks,and the real data can avoid large-area multi-path noise points.The error of the simulation data using the stable region and the utilization of all data registration are 0.0075 m and 0.0093 m,respectively,and the actual data are 0.0128 m and 0.0159 m,respectively,and the registration accuracy is improved.The magnitude and direction information of the deformation can be obtained by using the projection distance deformation of the nearest-point Euclidean distance in the reference data method.