Analysis And Application Of Typical Feature Extraction Method Of 3D Laser Scanning Point Cloud

At present,three-dimensional laser scanning,which is an emerging technology,has been widely used in production activities due to its high-precision surface data collection capabilities,but there are a large number of redundant point clouds in the collected data.The performance of the software and hardware puts forward higher requirements,and at the same time,increases the post-data processing time and reduces efficiency.In addition,in some production activities,it is necessary to collect data from the same area multiple times,and use the collected multi-period data for registration to analyze the differences between the point clouds collected at different times to achieve their respective needs.Therefore,for this situation,this paper uses different data extraction methods to extract common features in point clouds,and uses the extracted features of the same name to carry out the coordinate system between time-series point clouds.At the same time,the accuracy evaluation and field Application,the main work is as follows:(1)Use different supervised learning methods to generate supervised classification models to extract rod-shaped objects in the scanned point cloud.The basic steps of supervised learning,issues to be considered and other important factors are introduced;then the classification principles of different supervised learning methods are explained.The classification models generated by different supervised classification principles were applied to classify the data,and the confusion matrix is used to evaluate the accuracy of the classification results.The results show that the weighted K nearest neighbor classification has the best classification effect on the point cloud data extracted from the selected classification model.(2)Extract the feature parts of the building using the geometric features of the neighborhood.Applying spherical neighborhood search,construct a spherical neighborhood set of feature points under the gradient search radius.Calculate the geometric characteristics of the feature points and compare the changes of the calculated values of the same features in the gradient neighborhood to determine the optimal search radius of the point cloud;Redundant data still exists in the feature point cloud obtained from the geometric features of the neighborhood,and the extraction results are re-extracted using Gaussian curvature.(3)Taking the data of the simulated school experimental site as an example,the three-dimensional coordinates of the rod-shaped object are extracted using hierarchical least squares fitting.The three-dimensional coordinates of the feature points in the feature point cloud of the building are extracted using the projection point density of different projection directions;based on the benchmark The three-dimensional coordinates of the rod-shaped objects and building feature points extracted from the data are converted using 4PCS for coordinate system and using point-to-plane ICP for accurate registration;the coordinate conversion parameters are used to register the verification point cloud and evaluate its accuracy.The results show that the plane accuracy and elevation accuracy of the verified point cloud data after the coordinate system are 0.025 m and 0.015 m,respectively.In practical application,the average deviation between the measured subsidence obtained in the same area of the subsidence area and the predicted value is about 0.1m,and the trend has maintained good consistency.The paper has 74 pictures,10 tables,and 149 references in this thesis.