Semantic Segmentation Of 3D Point Cloud Based On Spatial Eight-Quadrant Kernel Convolution Network

Unmanned vehicles and intelligent automatic robots have always been popular technologies in scientific research and industrial production,aiming to achieve a better life.The application of these technologies requires the agent to be able to perceive the surrounding environment in real time and analyze the situation of various objects in the environment,including the type,location,and size of the object.It is hard for traditional 2D images to show the information of depth and distance.Instead,the 3D point cloud can provide a good data foundation for the environment perception of these agents.Focused on the study of semantic segmentation in 3D point clouds,this paper has made innovative work mainly in two aspects.First,a spatial eight-quadrant kernel convolution algorithm for three-dimensional point cloud feature extraction is proposed.Second,the proposed spatial eight-quadrant kernel convolution algorithm is embedded into some classic 3D point cloud semantic segmentation network,to improve the accuracy of the model's semantic understanding of the real 3D scene.First of all,by modeling the relationship between points in 3D space,a spatial eight-quadrant kernel convolution algorithm is proposed to solve the problems of common point cloud feature extraction methods,such as the loss of point cloud feature information and insufficient semantic expression.The algorithm we proposed is used to extract the semantic features of input point cloud,enrich point cloud feature information and effectively improve the semantic expression ability of features.Taking into account the disorder of point cloud,the algorithm first divides the local neighborhood of the point cloud space,then,extracts the features by grouping the point clouds and aggregating them.That is,first extract the point features in each small neighborhood in the point cloud space,and then sequentially expand the neighborhood range to continue to extract the point cloud features until the entire 3D point cloud space is covered.Taking into account the relevance between points in the point cloud,the algorithm uses spatial kernel convolution to perform convolution operations with K spatial neighbors in the process of neighborhood spatial feature extraction,so as to achieve the modeling of the point-to-point relationship.Considering the sparsity of the point cloud,the extracted features are further weighted by the inverse square distance,to improve the robustness of the model for feature extraction in regions with different degrees of density.Finally,the point cloud feature extraction algorithm proposed in this paper can fully extract the multi-scale spatial features of the point cloud from the local neighborhood of the point cloud.Therefore,it can effectively improve the problems of information loss and insufficient semantic expression when extracting point cloud features with existing methods.Secondly,the point cloud feature extraction of PointNet++,PointSIFT and PointConv network is not precise enough,more detailed information is lost in the process of feature propagation,and the extracted feature lacks multi-scale semantic information.Thus,a more effective point cloud down-sampling and up-sampling module is proposed to improve these three classic point cloud semantic segmentation networks.In the PointNet++ network,by directly replacing the original down-sampling feature extraction module and transforming the up-sampling module,the 3D point cloud semantic segmentation capability of the improved model has been promoted to a certain extent.In the PointSIFT network,the PointSIFT module of the network is retained,and the original set abstraction module is replaced with the proposed downsampling module.The semantic segmentation capability of the network is mostly enhanced.When the original corresponding module of PointSIFT network is replaced with the proposed up-sampling module,the network segmentation of some classes in the dataset is improved.In the PointConv network,when the proposed down-sampling module and the PointConv feature encoding module are used at the same time,the semantic segmentation results of the network have been further improved.At last,the performance of the proposed algorithm is verified on a large number of public datasets.All the results show that the point cloud features extracted by the spatial eight-quadrant kernel convolution algorithm can make up for the shortcomings of the existing partial network point cloud feature extraction and poor semantic information expression ability.At the same time,the proposed algorithm can be effectively embedded in different models to further improve the performance of the model.