Research On Point Cloud Segmentation Algorithm Based On Deep Learning

With the popularity of depth sensors and 3D scanners,3D point cloud has developed rapidly.3D scene understanding based on deep learning has become a research hotspot.The point cloud data processing of 3D scene includes object classification,object detection,instance segmentation and semantic segmentation,etc.Among these tasks,object classification and scene semantic segmentation are the hot topics of current research.However,there still exists some challenges in processing point cloud data.Due to the unstructured and unordered property of points,it is difficult for researchers to directly capture the complex relationship between points.Some networks only consider the original coordinates and the feature information of a single point,and do not pay enough attention to the local geometric relationship and feature acquisition of the point cloud,thus ignoring a lot of spatial geometric information.For the sake of enriching the feature representation of points and capturing much more contextual features,two methods are proposed to enhance the feature representation of 3D point cloud in our paper,which includes:1)A deep convolution network for object classification and semantic segmentation of point cloud is proposed.The encoding unit is designed to encode the information of each point in eight directions and the original coordinates and relative coordinates of local points,so that before sampling and grouping operation,the features of each point contain the features of its neighboring points;then,the local area of the point is obtained by grouping through the farthest sampling method and ball query method,and the features of local points are enriched by constructing the Graph Attention module.Finally,the pointnet network is applied to extract the local features,so that the final fusion of the center features contains not only its own original features,but also the feature information of the surrounding points,which is conducive to learning the complex local structure features of the point cloud;in addition,the network also proposes a new multi-dimensional loss function,which combines the classical cross entropy loss function with the face recognition.The central loss function is combined to act on the classification task.2)A classification and segmentation network architecture for learning global and local feature representation of point cloud is proposed.For the sake of solving the problem of insufficient expression of initial features of point cloud,a Data Augmentation Module is proposed in the network.For each input point,we take the features of neighboring points as its contextual information to enrich the semantic representation of points.We also design the Feature Convolution layer,and query the neighboring points of each point through the k-nearest neighboring search method to form the local regions.For each local region,the nearest neighbor digraph is constructed to represent the local structure of the point cloud,and the geometric relationship between the changing points is dynamically captured by generating edge features and direction features.For the local regions formed by sampling and grouping operation,a Local Feature Augmentation module is designed to enrich the feature expression of local points.A lot of experiments have been done on Model Net40,Scan Net and S3 DIS datasets.The experimental results show that our network has excellent performance in both classification and semantic segmentation tasks.We can apply our model to some real scenes such as interior scene design and layout analysis.