Research On Classification Algorithm Of 3D Point Cloud In Deep Learning

With the development of 3D sensor technology,how to effectively process 3D point cloud data has become a new research hotspot.The point cloud is a sampling of objects.It uses three-dimensional coordinates to describe the geometry and position information of objects.Point cloud classification algorithms are widely used in fields such as autonomous driving,virtual reality,augmented reality,3D face,SLAM,and remote sensing mapping.Among them,autonomous driving vehicles often use LiDAR as a sensor to obtain surrounding point cloud data.The point cloud classification algorithm gives the semantic information of the original point cloud to help the adjustment of vehicle speed and the choice of driving direction.In recent years,point cloud data classification technology based on deep learning has made considerable progress in simulating point cloud data,but it is still a challenging problem in the classification of real point cloud data with noise and background.First,this article discusses how to combine the Siamese neural network with the point cloud classification network PointNet.The point cloud classification framework based on the Siamese neural network requires two point cloud data to be input at the same time.For two point clouds input at the same time,the network uses the point-wise multilayer perceptron with shared weights to extract their local features.Then use the max pooling function to extract the global information and get the global features of the two point clouds.The model uses a feature reorganization function to process the global features of two point clouds.For point clouds with the same category,the feature reorganization function will retain features with similar activation values and attenuate the activation value features with larger differences.For point clouds with different categories,the feature reorganization function will retain the features with different activation values,and the activation value features with smaller attenuation differences.Finally,a fully connected network is used to implement the classification task.Secondly,this paper solves the problems of the point cloud classification framework based on the Siamese neural network and proposes a dual-input network framework.The dual-input network is similar to the Siamese neural network framework in that it also requires the input of two point cloud data at the same time.The dual-input network framework uses a neural network with feature extraction capabilities to extract point cloud information and uses a neural network with feature analysis capabilities to analyze the features of the point cloud.It uses a feature comparator to process the feature vector between the feature extractor and the feature analyzer.The feature comparator realizes the clustering of the point clouds of the same category in the highdimensional space.It narrows the distance between the point clouds with the same label and widens the distance between the point clouds with different labels.In the application of the dual-input network framework,this article respectively introduces the application and results of the PointNet network and PointCNN in the dual-input network framework.Finally,this paper compares the similarities and differences between the aggregation function and the TripletLoss loss function in FaceNet.Theoretical analysis shows that the aggregation function is more suitable for the dualinput network framework for implementing classification tasks.The experimental results of DI-PointNet and DI-PointCNN with dual-input network framework in simulated and real data sets show that the dual-input network framework improves the robustness of the original network,excludes some background noise interference,and is suitable for real Point cloud data.The average accuracy and overall accuracy of DI-PointCNN in the publicly simulated point cloud dataset ModelNet40 are 88.3% and 92.1%,respectively,and the average accuracy and overall accuracy of the public real point cloud dataset ScanObjectNN are 79.6% and 81.3%,respectively.