Research On Point Cloud Classification Method Of Large Scene Based On Residual And Self Attention

With the development of science and technology,large scene 3D point cloud is applied in more and more fields.Researchers continue to deepen the research on large scene 3D point cloud classification,and various classification models emerge in endlessly,and have made good achievements.Especially in recent years,the emergence of convolutional neural network has greatly promoted the progress and development of large scene 3D point cloud classification 3D urban modeling,unmanned ship cruise and other fields have been widely used.This paper proposes two different frameworks for 3D point cloud classification in large scenes.The specific work is as follows:(1)Aiming at the low classification accuracy caused by gradient disappearance /explosion and feature loss in convolutional neural network,a feature rnet based on residual convolutional neural network is proposed to realize point cloud classification.The framework model does not directly take the three-dimensional point cloud data as the input,but takes the two-dimensional features of the three-dimensional point cloud extracted by KNN method and the feature image constructed by the three-dimensional features as the input,which avoids the non adaptability of the two-dimensional network framework to the direct processing of the three-dimensional point cloud data and makes the three-dimensional point cloud suitable for the two-dimensional network;The rnet framework structure designed by the model uses the residual module.On the one hand,it can prevent the problems such as gradient explosion and network degradation.On the other hand,through the addition operation,the extracted deep features and shallow features are added to reduce the impact of information loss in the convolution process on the final result to a certain extent,so as to improve the classification accuracy.The open Oakland 3D data set and GML data set are used to train the feature rnet framework model respectively.The experimental results show that on the Oakland 3D dataset,compared with other existing deep learning classification frameworks,the proposed framework has a great improvement in classification accuracy,and the overall classification accuracy can reach 97.7%.At the same time,the feature rnet framework has also achieved good results on GML data sets.(2)Aiming at the problem of low classification efficiency and large amount of parameters,a lightweight convolutional neural network(light botnet)based on self attention mechanism is proposed.The framework is to extract features from point cloud data,construct feature images from adjacent feature points as the input of light botnet network framework,and use the model to complete the cloud classification of large scenic spots.The botnet framework applied to image classification is successfully applied to the task of three-dimensional point cloud classification.Experimental results show that compared with most existing point cloud classification methods,this method achieves an overall classification accuracy of 98.1% on Oakland 3D data set.At the same time,it reduces the number of parameters and training time as much as possible,which makes the network converge faster and reduce the amount of calculation.