Research On 3D Point Cloud Classification And Semantic Segmentation Algorithm Based On Attention Mechanism

Point cloud technology has received more and more attention in recent years,and it has a wide range of applications in computer vision,robotics,and autonomous driving.As an important technology in the field of artificial intelligence,deep learning has been successfully applied to solve various two-dimensional vision problems.In recent years,many research works have begun to use deep learning technology to process 3D point clouds,mainly involving point cloud classification,segmentation,target detection,3D reconstruction,point cloud completion and other applications.Among them,the classification and segmentation of point cloud,as an important basis for the research of point cloud technology,can be respectively applied to urban planning and construction and automatic driving,but the current point cloud classification and segmentation lack of interaction between points,resulting in the loss of information acquisition.Since the attention mechanism can effectively capture long dependencies between inputs,it is very suitable for interactive processing of point cloud information.Therefore,this thesis introduces the attention mechanism into the point cloud classification and segmentation tasks.The main research work is as follows:(1)Combined with local feature extraction,self-attention mechanism and multiscale feature fusion technology,a multi-scale point cloud classification neural network model based on attention mechanism is proposed.The number of input points in each layer of the network is half of the points in the previous layer,and the farthest point sampling(Farthest Point Sampling,FPS)is used to keep the point cloud shape at each scale as much as possible;for each scale,K nearest neighbor(K Nearest Neighbor,KNN)algorithm is used to obtain the adjacent points of each point to realize the extraction of local features,thereby enriching the features of each center point;then use the selfattention mechanism to perform feature interaction between all points at each scale to achieve global feature extraction;finally,a multi-scale fusion module is designed to automatically learn the weights at each scale to achieve feature fusion at different scales as the final feature.Experiments are conducted on public datasets,and the results show that the proposed network effectively improves the accuracy of point cloud classification.(2)On the basis of the point cloud classification model,this thesis further improves the attention mechanism and proposes a multi-scale point cloud semantic segmentation neural network model based on the attention mechanism by combining the theories of upsampling and multi-scale fusion.Specifically,the thesis proposes a position information extraction module based on synthetic attention mechanism and a semantic information extraction module based on self-attention mechanism from two aspects of position information and semantic information.Different from point cloud classification,semantic segmentation needs to pay attention to the features of each point.Therefore,the thesis uses interpolation to gradually recover the number of points in the upsampling process of the network,and uses skip connections to enrich the features of each point.Finally,the final features are obtained by utilizing the adaptability of different dimensions of each point through the channel attention mechanism.Experiments are conducted on public datasets,and the results show that the proposed network effectively improves the accuracy of point cloud semantic segmentation.