Research On Point Cloud Object Classification Based On Deep Learning

Point cloud is an important data formats in computer vision,which used in autonomous driving,robot,virtual reality and other fields.With the enhancement of lidar performance,the acquisition of point cloud data has become easier.And it put forward higher requirements for point cloud data processing.Object classification is an important branch of point cloud research.It is the basis of other tasks.Currently,Many point cloud classification methods based on deep learning fail to combine global features and local features,resulting the accuracy of network is not high.On the other hand,the device is easily restricted by the surrounding environment when collecting point cloud data,which causes various types of corruptions.These corruptions bring greater challenges to point cloud.In response to the above problems,this paper conducts deeper research on classification tasks,which the specific work is as follows:(1)Research multi-scale point cloud classification networks that combine global features and local features.First in the sample part,we use a global sample method to output a certain number of point clouds.And match these data with the initial point clouds.The output point clouds are important data that maintain the geometric shape and contain global feature information.Secondly,referring to the method of graph convolutional neural network,we design a local directed graph centered on the output points to learn local feature information.Then extract local features like graph convolutional neural network.Finally,we design multi-scale point cloud classification network Point MSS which use the above two methods.The experiment results show that the overall accuracy of Point MSS on the Model Net40 and Scan Object NN datasets reach 93.6% and 80.7%.We also reduce the number of sample points to verify the effectiveness of the global sampling method.The results show that the decrease speed of accuracy is slow,which proves the method has good performance.(2)Research data augmentation methods to deal with point cloud corruption.Classification networks usually focus on improving the effect of feature extraction,but this could not deal with data corruption.This paper uses data augmentation method to solve this problem.We design two kinds of point cloud augmentation methods.The first method is the local transformation of the point cloud.It divides point cloud data into several parts.At each part,use the combination transformation of rotation,scaling and translation.Finally adjust the transformed data by kernel regression function.The second method is the mix of point clouds.We use EMD to define the minimum distance measure of point clouds,and use the data in one sample replace the data of another sample to achieve smooth mix on different data.We verify these augmentation methods on two point cloud corruption datasets,Model Net40-C and Model Net-C.The experiment results show that the transformation augmentation performs well on deformation corruption.The mix augmentation performs well on noise corruption.It indicates that data augmentation of point cloud can effectively improve the robustness of the network.