Research On Point Cloud Classification And Semantic Segmentation Technology Based On Deep Neural Network

With the rapid development of 3D imaging technology,the cost of acquiring 3D data is getting lower and lower.The emergence of more and more large-scale 3D data sets has laid the foundation for the application of deep learning to 3D data,and the practical significance of3 D data has attracted a large number of scholars to invest in 3D deep learning research,and has promoted many applications that rely on 3D point cloud data,such as autonomous driving,robotic manipulation,and virtual reality.The common research tasks in these applications are 3D point cloud object classification and 3D point cloud semantic segmentation.Therefore,research on 3D point cloud object classification and 3D point cloud semantic segmentation has important scientific research and practical application value.Aiming at the situation that existing 3D point cloud classification and segmentation algorithms only consider a single neighborhood,this paper proposes a Dilated K-nearest neighbor feature enhancement module,which consists of a dilated K-nearest neighbor module,a point cloud neighborhood fusion module and a multi-scale neighborhood feature fusion module.By drawing on the idea of dilated convolution in twodimensional images to increase the pixel receptive field,the Dilated Knearest neighbor module selects the dilated neighborhood point cloud of the center 3D point by jumping,which greatly increases the receptive field of the center 3D point.The size of the receptive field allows the central 3D point to have feature information of a wider range of neighborhoods.The point cloud neighborhood fusion module uses the attention mechanism to fuse the feature information of the central 3D point itself and the dilated neighborhood feature information to enhance the feature information of the central 3D point,the multi-scale neighborhood feature fusion module fuses the multi-scale feature information of the center 3D point,and further enhances the feature information of the center 3D point.In addition,in order to avoid discarding feature-rich 3D points in the point cloud downsampling process,this paper proposes a feature-based sampling module,which to a certain extent avoids the selection of 3D point cloud subsets based solely on the distance information of 3D points.Instead,the size of the feature information of the 3D points is used to select a suitable 3D point cloud subset.Even if downsampling is performed,too much feature information will not be lost,thereby enhancing the accuracy of the 3D point cloud classification task and the semantic segmentation task.This paper is validated on the Model Net40 classification dataset and the S3 DIS segmentation dataset,and the experimental results show that the proposed algorithm model can achieve better performance in representative public datasets for point cloud classification and segmentation tasks compared with other state-of-the-art methods.This paper includes 41 images,7 tables and 61 reference doucumentations.