Study On LiDAR Data Classification Based On Convolutional Neural Network

The Light Detection and Ranging(LiDAR)system is an effective technology for obtaining high-resolution data.The acquired data is widely used in important fields such as topographic mapping and urban planning.One of the basic tasks in many fields is the classification of features.However,the existing feature extraction methods mainly focus on the characteristics of a certain aspect of the feature,and use the linear or nonlinear equation to artificially specify the extracted features and it is difficult to adjust the parameters.For deep learning,it allows the computer to automatically learn the features that are beneficial to the classification task and incorporate the process into a part of the model training,which helps to further improve the classification recognition accuracy.Therefore,this paper focuses on the research and improvement of LiDAR data based on convolutional neural networks.The main contents of this paper include as follows:Firstly,the development history of LiDAR data characteristics is reviewed.Several typical feature extraction and classification methods for LiDAR data are introduced.The differences and usage scenarios between different methods are given,and the evaluation criteria of classification results are described,and provide a theoretical basis for the subsequent chapters of this article.Secondly,the method of classifying LiDAR data based on convolutional neural network is designed and implemented.The nonlinear feature of powerful representational capabilities of LiDAR data is extracted based on convolutional neural network.The convolutional neural network is combined with the morphological operation and compared with the traditional support vector machine classifier method to verify the superiority of the method.On this basis,the new activation function SiLU is introduced after studying the various modules of the convolutional neural network,which reduces the problem of partial neuron inactivation and increases the self-stability of LiDAR data in the process of feature extraction.Finally,the LiDAR data classification method based on spatial transformation network is studied in detail.Because the traditional convolutional neural network is based on experience or experiment,the fixed rectangular window contains information about the surrounding objects,but the remote sensing objects have great differences in size,direction and size.Here,a spatial transformation network is used.The function of rotation,translation and scaling of the object target makes it more suitable for the processing of subsequent convolutional neural networks.In addition,this paper introduces the multi-attribute profile.Experiments show that the improved spatial transformation network has more robust feature extraction and classification capabilities.