Point Cloud Semantic Segmentation Method Combining Feature Selection And Deep Neural Network

Semantic segmentation technology is the basis of 3D scene point cloud data understanding and analysis,and has become the key research content in the fields of Surveying and Mapping Geographic Information Technology,Water Conservancy Engineering,Navigation and Positioning and Unmanned Driving.This paper proposes two algorithms to achieve efficient and accurate semantic segmentation of outdoor large-scale point clouds.This paper uses traditional feature extraction methods to construct multi-scale spherical neighborhoods to generate initial feature description vector(IFDV),and then designs a feature selection unit that includes a two-layer feature screening structure to perform features on the basis of IFDV to generate an optimal feature description vector(OFDV).The OFDV contains the optimal feature descriptor of point cloud data,which provides advanced prior semantic information for subsequent neural network learning.In the first semantic segmentation algorithm,this paper designs a point cloud semantic segmentation network based on optimal feature description vector(OFDVNet).The network takes the optimal feature description vector as input,then designs a neural network structure with 6-fully-connected layers,and finally realizes point cloud semantic segmentation.To verify the effectiveness and universality of the proposed network,this paper applies OFDV-Net to two standard public outdoor large-scale point cloud benchmark datasets SEMANTIC3D.NET and Paris-3D,achieving 88.3%and 87.7%overall segmentation accuracy respectively and average intersection ration of 57.3%and 47.5%.Moreover,the training time required for the proposed network is relatively short.Experiments on the outdoor large-scale point cloud benchmark datasets show that OFDV-Net can effectively shorten the model training time while obtaining high-precision semantic segmentation results for outdoor large-scale point clouds.This paper selects the point cloud data of river bank scene in the field of water conservancy engineering for verification,and the results show that ofdv net is also suitable for the point cloud data of water conservancy scene.In the second semantic segmentation algorithm,a point cloud semantic segmentation network based on optimal feature description matrix(OFDM-Net)is designed on the basis of OFDV-Net.By solving the K nearest neighbors of each 3D point,the network combines the k-nearest neighbors and the 3D points are combined and arranged to construct an optimal feature description matrix(OFDM).Using OFDM as input,a convolutional neural network model is designed to realize point cloud semantic segmentation.To verify the effectiveness and universality of the network,this paper applies OFDM-Net to the public outdoor large-scale point cloud benchmark dataset SEMANTIC3D.NET,achieving 90.1%and 59.7%of the overall segmentation accuracy and average intersection ration respectively.Compared with OFDV-Net,the overall segmentation accuracy and average intersection ratio are improved by 1.8%and 2.4%,respectively.