Research On Classification And Segmentation Method Of 3D Point Cloud Based On Deep Learning

Point cloud is a common spatial 3D data format.Currently,3D point cloud processing technologies,such as 3D point cloud object classification and semantic segmentation,have important research value and a wide range of application scenarios in the computer field.With the success of deep learning architecture,the methods of3 D point cloud object classification and semantic segmentation have shown impressive results.In deep learning-based methods,Point Net++ is the basic framework for processing point cloud data.Although the framework can achieve high classification and segmentation accuracy,it still encounters with some challenges:(1)When extracting features from the local region,Point Net++ simply encodes the geometric features by the relative coordinates between the neighbor point and center point.This is not enough to fully characterize the geometric shape from the local region,leading to incomplete information extraction for the input point cloud.(2)Poine Net++ ignores the learning of the relationship between points and fails to explore the spatial distribution of local regions,which severely limits the representation ability and reasoning performance.(3)Due to the high computational complexity of the farthest point sampling in Point Net++,it is difficult to be applied to mobile terminals,which limits the application of this method in real scenarios.To address the above challenges,this paper carried out the following research work:(1)This paper designs a point cloud classification and segmentation network based on multi-information feature enhancement,named FE-Point Net++.Specifically,on the basis of using relative coordinates in Point Net++,FE-Point Net++ further adds the center point coordinates,neighbor point coordinates,and the euclidean distance between these coordinates.In this way,the network can more effectively learn the geometric shape information of the space.Moreover,coordinate information and additional information are split and encoded to maximize the input information.Experimental results show that the overall accuracy index of the proposed method is92.3% on the Model Net40 dataset,and the m Io U index is 55.8% on the S3 DIS dataset.(2)Combined with the FE-Point Net++ network,this paper presents a space-aware point cloud classification and segmentation network SAFE-Point Net++.Specifically,a spatial awareness module is designed to fuse spatial distribution features and input features.With the fused feature,the module further models the relevance between the weight of each point and all points in the local region.This helps to explicitly perceive the local fine-grained shape and improves the accuracy and robustness of classification and segmentation.Then,the residual connection is assigned to prevent the network model from overfitting.Experimental results shown that the overall accuracy index of the proposed method is 93.4% on the Model Net40 dataset,and the m Io U index is 58.1%on the S3 DIS dataset.(3)Furthermore,this paper migrates the SAFE-Point Net++ to the embedded edge computing platform Jetson TX2.In particular,a feature-based sampling method is designed to replace the farthest point sampling algorithm in SAFE-Point Net++.It can reduce the computational cost for the point cloud sampling and ensure classification and segmentation accuracy.Then,the improved network is ported to Jetson TX2 by Tensor RT acceleration.Experiments on Model Net40 show that the reasoning time of the improved network is not further increased compared with the server side.More importantly,it still achieves a high classification accuracy.This shows that the proposed algorithm has practical applications in mobile terminals.