Research On Automatic Recognition Of Urban Objects Based On Deep Learnin

With the rapid development of economy and society,China’s urban landscape has undergone tremendous changes.At present,China’s surveying and mapping geographic information industry is vigorously carrying out the construction of digital cities,smart cities and real three-dimensional China.In the process of these project construction,urban feature recognition is an important work.How to effectively and accurately identify urban feature category information is of great significance for the construction of these projects.Airborne Li DAR and UAV oblique photogrammetry are the main technical means to obtain spatial information of urban features,which can provide basic data for urban feature recognition.Based on Li DAR and oblique photogrammetry data,when using deep learning models for urban feature recognition,existing network models fail to make full use of effective information,resulting in low accuracy of urban feature category recognition.In addition,when the amount of sample data is small,it will result in low recognition accuracy and insufficient model generalization capacity.Therefore,based on the theory and algorithm of deep learning,this paper uses the point cloud data obtained by Li DAR and oblique photogrammetry to preprocess the data to obtain LG dataset and combination strategy dataset.Then,through the constructed convolutional neural network model,the urban feature information is extracted and compared with other conventional models,and the automatic recognition of urban features based on deep learning is carried out.The main research contents and conclusions of this paper are summarized as follows :(1)The paper first introduces the basic unit of neural network based on deep learning.Then,the parameter training process of the network model is analyzed.Finally,the basic principle of the convolutional neural network model based on point features is summarized,which provides a theoretical basis for automatic recognition of point cloud urban features.(2)Aiming at the problem that the Point Net model cannot effectively improve the recognition accuracy of dense ground objects when extracting geometric features and location information of point clouds,this paper constructs an improved Point Net model that integrates a multi-head attention mechanism.The improved model is used to identify urban features.Firstly,in the Point Net network model,a self-feature enhancement module is introduced to fuse feature correlation between points to obtain context information and global feature information.Secondly,through the point cloud structure dimension,the SE-Net module is improved to obtain the SE-Point improvement module suitable for point cloud data.Then,the SE-Point improvement module is added to the local network T-Net,and the local network TNet is improved by using the relationship between point cloud channels,thus improving the model’s ability to focus on effective information.Finally,the deep convolutional neural network encoder is used to extract the feature information of the point cloud.Based on the LG dataset,this paper carried out experimental research on urban feature recognition and compared it with conventional Point Net,Point Net++,Rand LA-Net and SCF-Net models.The experimental results show that compared with the conventional five comparison models,the improved SA-Point Net model constructed in this paper has higher accuracy and is more conducive to improving the accuracy of urban feature recognition.Its overall accuracy,recall rate and m Io U reached 92.24%,86.78% and 75.56% respectively.(3)Aiming at the problem of low model accuracy and insufficient generalization ability caused by less sample data when using deep learning algorithms for urban feature recognition,this paper carried out point cloud data augmentation research.When using geospatial data for data augmentation,first,modeling data is preprocessed.Then,the modeling scene is imported onto the Airsim platform to obtain two-dimensional image data.Finally,using the image data,the point cloud data is generated by the dense matching algorithm.Based on the augmented combination strategy dataset,this paper carried out experimental research on urban feature recognition.The improved SA-Point Net model constructed on this paper is compared with the conventional Point Net,Point Net++,Rand LA-Net and SCF-Net models.The analysis results show that the accuracy,recall and m Io U of the improved SA-Point Net model are 88.49%,87.13% and 77.91%,respectively,and the overall accuracy of the model is 93.62%.Compared to the LG dataset,the overall accuracy is higher when using the augmented combination strategy dataset for feature recognition.(4)Finally,based on the improved SA-Point Net model and the data augmentation combination strategy dataset,the paper uses UAV tilt photogrammetry point cloud data to carry out the case study of automatic urban ground object recognition,and verifies the feasibility and effectiveness of the ground object recognition model constructed in this paper.The research shows that based on the LG dataset,the overall accuracy of the improved SAPoint Net model for urban feature recognition is 88.79%,and the improved model is better than the Point Net model.Based on the combined strategy dataset corresponding to data augmentation,the overall accuracy of ground object recognition of the improved SA-Point Net model is 90.90%.Compared with the LG dataset,the augmented combination strategy dataset has higher accuracy in urban ground object recognition,which verifies the feasibility and effectiveness of constructing the improved SA-Point Net model and the data augmented combination strategy dataset.