Research On Point Cloud Semantic Segmentation In Large-scale Urban Scenes Based On Self-supervised Learning

Due to the excellent spatial expression ability,point cloud has become an important research data in remote sensing,geographic information,computer vision and many other fields.Large-scale urban point cloud generated by airborne Li DAR(Light Detection And Ranging)system or oblique photography technology can express 3D surface information in a large range of research areas,and has extensive applications.Point cloud semantic segmentation is the basis of these applications,and is also the focus of research.In recent years,the rapid development of deep learning has led to the emergence of a large number of methods in point cloud semantic segmentation,most of which rely on labeled data and ignore the information contained in a large amount of unlabeled point cloud.Selfsupervised learning can construct target tasks based on data itself,which has played an important role in natural language processing and image fields.Therefore,this thesis introduces a self-supervised learning method,which will fully learn features with rich geometric information from unlabeled point cloud data based on point cloud reconstruction tasks,and thus bring assistance to the downstream task,i.e.,point cloud semantic segmentation.The main research work is as follows:(1)This thesis designs a geometric feature learning network based on self-supervised learning,which takes point cloud reconstruction as its target task and presents a discrete autoencoder structure,with the purpose of learning general geometric features of training scenes.The whole network takes the local patch of point cloud as the basic unit for feature learning.Therefore,this thesis uses the farthest point sampling,k nearest neigbhors and multi-layer perceptron for neighborhood construction and embedding to realize the geometric information encoding of the point cloud.The Edge Conv module forms the main part of the network to complete feature extraction,the Folding Net module mainly realizes the 3D coordinate mapping of potential features to the reconstructed point cloud,and the dictionary between the encoder and decoder is mainly used to summarize the geometric details in the training scene.The visualization and loss results obtained from the point cloud reconstruction experiment based on DALES dataset show that the model can effectively complete the point cloud reconstruction task.In addition,experiments on the influence of parameters such as sphere radius,sample size,and dictionary size have shown that reasonable settings can lead to lower reconstruction losses and better reconstruction results.(2)For the purpose of verifying the effectiveness of the geometric features learned by the self-supervised model,the point cloud semantic segmentation network is designed in this thesis.On the one hand,the network includes neighborhood embedding,encoder and dictionary in the geometric feature learning network;on the other hand,it presents a hierarchical network architecture similar to Point Net++ to realize multi-level and multiscale feature extraction.The geometric features output by the former will be concatenated into the hierarchical network architecture as auxiliary information.The semantic segmentation results on the DALES dataset have preliminarily verified the effectiveness of the proposed method,with OA and m Io U reaching 92.5% and 72.3%,respectively.Compared with the baseline method,OA and m Io U have increased by 1.1% and 3.8%,respectively,indicating that the learned geometric features can play a role in downstream tasks,resulting in improved accuracy.The comparison results with other deep learning methods are good,with m Io U ranking second among all methods.In addition,the proposed method can also achieve good experimental results on the Swiss dataset,with OA and m Io U reaching 93.6% and 68.0%,respectively,and the comparison with the baseline method further verifies the effectiveness of the learned geometric features.