Research On Semantic Segmentation Method Of Urban Point Cloud Based On Multi-scale Feature And Graph Convolution Network

Typical urban features include buildings,roads,trees,etc.The accurate extraction of typical urban features plays an important role and significance in urban development planning,management,and the construction of smart cities.With the rapid development of technologies such as Li DAR and multi-view image matching,the acquisition of 3D data is becoming increasingly convenient.The urgent need to achieve rapid and accurate 3D precise extraction for urban scenes in order to meet the pressing demand for high-quality,refined 3D geographic information in smart cities has become a critical problem that needs to be addressed.In contrast to simple indoor three-dimensional scenes,urban three-dimensional space has high complexity,interlacing,and dynamics.Urban point cloud data faces problems such as large differences in terrain scales and complex and diverse structures.Traditional point cloud segmentation methods based on artificial feature design suffer from tedious processing and excessive reliance on transitions.In recent years,data-driven deep learning technology has achieved certain results in 3D point cloud semantic segmentation with its powerful feature learning capabilities.This thesis is based on the lightweight point cloud semantic segmentation network Rand LA-Net,this thesis focuses on the two aspects of multi-scale feature fusion and local feature enhancement.The main research contents of this thesis are as follows:(1)Semantic segmentation of urban point clouds based on multi-scale feature fusion.Aiming at the problem that the scales of different target objects in urban scenes are quite different,based on the Rand LA-Net point cloud semantic segmentation network,a point cloud semantic segmentation network Rand LA-MSFF(Rand LA-Net based on Multi-Scale Features Fusion),which fuses the cross-scale features of the point cloud based on the multi-path skip connection between the encoder and the decoder,and at the same time fuses the multi-scale features by upsampling inside the decoder;Aiming at the problem that the multi-path long skip connection between the encoding end and the decoding end of the Rand LA-MSFF network is prone to cause semantic gaps.Based on the Rand LA-MSFF network,a short densely connected multi-scale feature fusion point cloud semantic segmentation network Rand LA-MSFF++ is proposed in the encoder and decoder.Experimental results show that the proposed multi-scale feature fusion point cloud semantic segmentation network can better utilize spatial detail information and semantic features,and retain more neighborhood information.Taking the Rand LA-MSFF++ network as an example,the dense connection of the network makes up for the problem of semantic gap,and the features are progressive and gradually changed layer by layer,so as to better understand the deep feature information and achieve higher semantic segmentation accuracy.The m Io U score is improved by 3.12% and 4.6% compared with the backbone network Rand LA-Net on the two data sets of airborne radar point cloud NJSeg-3D and public photogrammetry point cloud Sensat Urban,respectively.(2)Embedding Graph Convolution for Semantic Segmentation of Urban Point Clouds.Aiming at the problem that the multi-scale point cloud semantic segmentation network is poor in recognizing complex structures in urban scenes,the embedded graph convolution is proposed to better obtain the semantic features of the point cloud geometric relationship and high-dimensional feature space,and improve the local feature extraction ability;For the edge of the two types of ground objects,the extracted features have the problem of mutual interference.The self-attention mechanism is used to dynamically learn the spatial position characteristics of the point cloud and the graph convolution to obtain high-dimensional semantic features.That is,the positional relationship and semantic information of the local neighborhood points are considered at the same time,so as to better realize the feature recognition on the edge of the feature.In order to further increase the receptive field of graph convolution,a graph convolution with enhanced receptive field is proposed to improve the semantic segmentation ability of target objects in urban scenes.The experimental results show that the multi-scale point cloud semantic segmentation network GCN-Rand LA-MSFF++ after the embedded graph convolution has improved the recognition of buildings with complex structures,non-trunk roads,etc.,and the edges of ground objects.It further reduces the problems of missed detection and false detection of typical urban features.The average intersection and union ratio m Io U of the GCN-Rand LA-MSFF++ network is higher than that of the backbone network Rand LA-Net in the lidar point cloud NJSeg-3D and photogrammetry point cloud Sensat Urban.The data sets have been improved by 3.86%and 6.7% respectively.This thesis has 36 figures,16 tables and 99 references.