Research On Combined Filtering And Feature Contour Extraction Based On Building Point Clou

The three-dimensional contour information of buildings holds significant importance for the development of digital cities,and building point clouds have become an essential data source in this application.However,the original point clouds are scattered,unordered,and characterized by large data volumes and severe noise interference,which pose challenges to point cloud filtering and three-dimensional contour extraction techniques.Therefore,in this study,in-depth research was conducted on point cloud filtering and three-dimensional contour extraction algorithms.The specific details are as follows:An improved dynamic statistical filtering algorithm was proposed to address the issue of feature information loss due to the neglect of point cloud distribution in traditional statistical filtering methods.The dynamic threshold based on neighboring point relationships was added to the global threshold of statistical filtering.To tackle the problem of different types of noise points that cannot be handled by a single filtering algorithm,an analysis of the causes and classifications of noise generation was conducted.Based on the improved dynamic statistical filtering mentioned earlier,a combined filtering strategy was designed.Firstly,the improved dynamic statistical filtering was used to remove distant noise points and reduce time complexity.Finally,radius filtering was applied to eliminate noise points near the buildings.Through experimental comparison analysis,it was found that the dynamic statistical filtering effectively reduced the filtering time when the point count exceeded fifty thousand.Meanwhile,the combined filtering method proposed in this paper achieved lower Type I error while ensuring a lower Type II error,effectively removing various types of noise points in buildings.An implicit representation-based unsigned convolutional occupancy network,UCONet,was designed to address the problem of simple fully connected structures being unable to process large-scale point clouds of buildings for obtaining their 3D contours.The designed convolutional encoder and implicit occupancy decoder were combined to aggregate global and local information of the point cloud and decode occupancy values of 0 or 1 for sampled points.To overcome the problem of implicit representation-based3 D contour extraction algorithms requiring normal information,an unsigned crossentropy loss function,UCE,was designed to optimize the UC-ONet network.Finally,a multi-resolution extraction method was used to extract the iso-surface with an occupancy value of 0.5 as the final 3D contour result.Experimental results showed that the proposed method preserved the rich details of building 3D contours while having better scalability for huge scenes and better robustness for more complex topological structures of building point clouds.