Research On 3D Point Cloud Semantic Segmentation Method Based On Deep Learning

3D point cloud is a common data format in the field of computer vision.Because the point cloud data contains the depth information of the scene,it has a wider application space than image.In recent years,with the rapid development of 3D scanning equipment,the acquisition of 3D point cloud data has become more convenient and cheaper.The fine analysis of point cloud data is one of the important means of robot environment perception.It is a research hotspot in the fields of automatic driving,robot,industrial detection and so on.However,feature operators designed by traditional methods are difficult to cope with increasingly complex application scenarios,while data-driven deep learning methods have made great progress in the fine analysis of point clouds.This paper focuses on the semantic segmentation of 3D point cloud,designs and implements the semantic segmentation algorithm suitable for indoor and outdoor scenes,including the following contents:1.Aiming at the problem of how to extract effective semantic segmentation features from large-scale real scene scenic spot cloud data,a semantic segmentation method of point cloud data based on bilateral structure is proposed.The network makes full use of the feature information of point cloud by using the geometric and semantic features of point cloud respectively.Aiming at the problem of excessive information loss in the commonly used maximum pooling method,a hybrid pooling structure is constructed to reduce information loss.Then,the attention mechanism is used to extract global features in two dimensions of space and channel,filter out scale noise and enhance the spatial expressiveness of features.Experimental results show that the method described in this paper can effectively improve the accuracy of point cloud semantic segmentation in different application scenarios.2.Aiming at the problem that the existing convolution neural network structure in the field of image processing can not be directly used for irregularly distributed point cloud data,a convolution method which can be applied to point cloud data is proposed.Firstly,the weight of each point is extracted according to the multidimensional measurement relationship between different points in the local area,and the convolution operator is constructed.The convolution neural network structure based on regular grid is extended to the point cloud data in the form of irregular grid.By introducing the volume ratio,the tolerance of the network to the geometric deformation of the same kind of objects is enhanced.Experiments show that the proposed method can effectively apply convolutional neural network to threedimensional point cloud data.The experimental results show that the method described in this paper can effectively extract point cloud features in different application scenarios,so this method can be extended to other fields that need to extract point cloud features,such as 3D reconstruction,slam and so on.