Research On Outdoor Scene Classification And Reconstruction Method Based On 3D Point Cloud

With the proposed of innovative technologies such as autonomous driving and digital cities,semantic segmentation of outdoor scenes based on 3D point cloud data plays an important role in the rapid development of applications such as analysis of autonomous driving road conditions and intelligent autonomous navigation.In terms of digital city construction,outdoor scene reconstruction technology has practical applications in areas such as urban planning,building restoration,and augmented reality.Semantic segmentation and reconstruction of outdoor scenes has attracted more and more attention from scholars,and has gradually become a research hotspot in the fields of computer vision and computer graphics.At present,there are still some problems in algorithm automation and complexity for the outdoor scene semantic segmentation of 3D point cloud.This paper focuses on the research of automatic semantic segmentation technology of the complex outdoor scene point cloud and the outdoor scene reconstruction method based on the semantic segmentation result.The main work is as follows:(1)This paper improves the DGCNN network and uses the convolutional neural network to complete the semantic segmentation of the scene point cloud.The improved convolutional neural network structure directly takes the original point cloud data as input,removes the alignment network,and uses EdgeConv in DGCNN to extract features.In the network structure,we reduce the number of convolutional layers and the amount of network parameters,making the network structure more lightweight.Then,we use the Huangshi public data set to verify the improved deep learning network in this paper.Finally,the semantic segmentation results of the network in this paper are visually compared with the results of DGCNN network.The improved method in this paper is evaluated in terms of overall accuracy and Intersection over Union(IoU).Experimental results show that the improved deep learning convolutional neural network model in this paper performs better than DGCNN network in outdoor scene semantic segmentation.In particular,the semantic segmentation performance in the category of poles and wires in outdoor scenes is superior to the DGCNN network model(2)Based on the results of semantic segmentation,this paper reconstructs the building walls,poles,wires and ground point clouds,respectively.According to the data characteristics and geometric characteristics of different objects in the scene,we use the "divide and conquer" idea to choose different methods to reconstruct different objects.In the reconstruction of building walls,region growing method based on voxelization is used to split the wall into a single plane Then the vertex coordinates of a single plane is calculated.Finally,the wall is reconstructed according to the vertex coordinates.In the reconstruction of poles,Euclidean distance clustering method is used to segment the pole.Then the height and endpoint coordinates of each pole are calculated.Finally,we use the cylinder to represent the pole.In the reconstruction of the wire point cloud,we also use Euclidean distance clustering to segment the wires.Then the single wire vertex and the length of the wire point cloud are calculated to complete the wire reconstruction In the reconstruction of the ground point cloud,based on the characteristics of uniform density distribution of ground point cloud data,the greedy projection triangulation algorithm is used to reconstruction the ground point cloudThis article takes the raw 3D point cloud data as the processing object.We discuss and study the outdoor scene classification method and the reconstruction method based on the semantic segmentation results.The use of convolutional neural networks to complete the semantic segmentation of outdoor scene point cloud plays a role in the development of point cloud data processing and scene reconstruction.