Research On Semantic Segmentation Method Of Airborne LiDAR Point Cloud Based On PointNet

With the continuous progress of intelligent robots,autonomous driving,smart city,virtual reality and other technologies,as well as the popularization of three-dimensional data acquisition devices and the rapid development of deep learning technology,semantic understanding and analysis of scenes have been continuously and profoundly developed.As the basis of scene understanding and analysis,point cloud semantic segmentation and 3D object classification related to computer vision have achieved many outstanding achievements.Three-dimensional information becomes an indispensable part of work.The airborne LiDAR(Light Detection And Ranging)system for capturing three-dimensional information becomes a new technical means for visualized spatial information because of its high precision,fast speed,strong reliability and low cost.At the same time,with the rise of deep learning in the field of computer vision,the method of directly applying deep learning to three-dimensional point clouds has attracted more and more attention.PointNet is a pioneering work that applies deep learning method to disordered point set and performs point cloud classification task.However,due to the structural characteristics of PointNet,its use of point cloud features is insufficient and the accuracy is not high.Aiming at this problem,a multi-scale and multi-feature algorithm based on the improved PointNet is proposed.(1)Multi-scale neighborhood design and multi-feature extraction of airborne point cloud data.Aiming at the shortcomings of the existing PointNet-based airborne LiDAR point cloud classification methods,this paper proposes a method to obtain the characteristics and hierarchical details of the point cloud by studying the spatial index structure of the point cloud and selecting the spherical neighborhood.According to different radius parameter combinations(R=0.8m and 1.2m),multi-scale features were extracted effectively,including basic features xyz.The RGB information obtained by the fusion of remote sensing image and point cloud and the total variance,roughness,flatness and other information calculated according to the eigenvalue are used to replace the single point as the input point of the network with the sampled neighborhood feature point set.(2)Build an improved PointNet neural network model.In this paper,PointNet's network architecture and parameter setting are used for reference to the training of the proposed model.Due to the addition of features,the dimension of the input transform part of the original PointNet was changed from 3 to 6,and three convolution layers were added in the feature extraction part.The 5-dimensional neighborhood features of the two sizes obtained above and the combination of xyz RGB are input into the mini-network.Due to the increase of data volume,L2 regularization was added to prevent overfitting and the dropout ratio was reduced.Data jitter and rotation were added in the data preprocessing part.Finally,the above point clouds combined with the multi-scale neighborhood,namely,the point clouds including the features under different radius parameters,are input into the improved network.In this form,effective features with strong generalization ability can be learned from the data.(3)In order to verify the effectiveness of the model proposed in this paper,this paper conducts horizontal self-comparison with single-scale single-feature data,single-scale multi-feature data and multi-scale multi-feature data respectively.The overall accuracy of multi-scale multi-feature data is 88.1%,which is the highest among the three combination methods.Longitudinal comparison was made with the experimental results of different research methods provided by ISPRS website,and the precision results of the proposed method were compared with some submitted results,and the classification accuracy was improved.In order to further evaluate the generalization ability of the proposed algorithm and the robustness of the network,the data of Sand City were processed with correlation,and the proposed algorithm was used to classify the data set.The accuracy of the classification results reached 92.4%.From the final results,it can be seen that the proposed multi-scale multi-feature method based on the improved PointNet can make full use of the point cloud features and improve the classification accuracy.