Research Of Point Cloud Classification Method Base On Convolutional Neural Networks

Point cloud classification is an important part of point cloud data procession.High-precision quick point cloud classification algorithm is very useful in auto-drive technology,robot technology,3D scene recognition,map reconstruction.In computer vision domain,convolutional neural networks(CNN)model arrives the state of art in difficult classification mission and surpasses all of the traditional algorithms which bases on the hand-extract feature.Because of the form of point cloud is not compatible with CNN,point cloud data can't be processed by existing CNN technology.But there is not any method allow us to process point cloud data with CNN technology.Base on the condition,this thesis has done follow work.CNN-base model is propose to process point cloud data.A special encoding method is used to transfer the point cloud data to matrix which allow the CNN processes the point cloud data.Depending on end to end training,this method can avoid the man-made bias furthest.Aiming at the problems of model,too many learnable parameters and low classification accuracy,this thesis uses the newly CNN and deep learning technology to improve the model.Learnable parameters are reduced by using the network construction principle base on SqueezeNet and deep compressed technology.Classification accuracy is improved by using the multi-scale information channel technology and residual network.The experiences show that construct CNN by the SqueezeNet principle can reduce the number of parameters without lost the accuracy.In addition,using the multi-scale information and res-connetion can improve the model performance.Verifying the method by implementing an engineering case.Using the CML Airbrne LIDAR Datasets as the experiment dataest,applying the method in chapter 2 and chapter 3 to classify points in point cloud.Verifying the method by analyzing and visualizing the experiment result.At the end,this thesis will discuss the shortcomings of this method and point out the possible improving way.At the end of the thesis,the drawbacks of the point cloud classification model based on convlutional neural networks are pointed out.Besides,methods and ideas for improving the point cloud classification model are pointed out so that the later researcher can refer to them.