Research On Airborne Point Cloud Classification Based On Deep Learning

Under the continuous development and promotion of modern science and technology,a series of emerging concepts such as “smart city”,“digital earth” and “digital city” are constantly being proposed.At the same time,with the maturity of 3D laser scanning technology,the collection and acquisition of large-scale cities data has become more simple and convenient,and the processing and classification of these data is a key step is to achieve intelligent urban analysis.Especially,the large-scale urban point cloud scanned by airborne laser contains a large number of object categories and many adjacent or overlapping parts,which brings great challenges to multi-objective classification.Aiming at this problem,this paper focuses on the main idea of multi-objective automatic classification and extraction of large-scale urban scenes,focusing on the following aspects:Firstly,the urban airborne point cloud is preprocessed,and then the 3D nearest neighbor is optimized based on the nearest neighbor space.For each point cloud,the 3D features of the nearest neighbor are solved.By analogy with the extraction method of 3D feature,airborne point clouds are projected onto 2D xoy plane,2D features are extracted based on 2D optimal circle neighborhood,feature selection is carried out using filter-based method and information gain strategy measure,and then 2D and 3D features are combined horizontally.The effectiveness of the proposed method is tested by using the constructed convolutional neural network.The experimental results show that the overall classification effect of the combination feature and the overall classification result of the 3D feature are 94.12% and 97.01%,respectively.Because of the sparse property of AlexNet model,it can fully learn the features related to training data.In order to further improve the classification efficiency of airborne laser point clouds,this paper improves and uses AlexNet to classify and recognize the combination feature matrix,and the overall accuracy can reach 97.79%.