Research On Large-Scale Point Cloud Classification Method Based On Multilayer Perceptron

The three-dimensional representation of urban scenery is the spatial positioning framework and analytical foundation of digital cities.In the practice of digital cities,point cloud data is one of the significant forms of three-dimensional geometric information.It is a key factor in achieving realistic3 D modeling and analysis of real-world scenes.Its practical applications are constantly expanding and deepening,particularly in fields such as autonomous driving,power line inspection,and forest monitoring.With the assistance of deep learning,the accuracy and efficiency of data processing have been continuously improving.Point cloud data is characterized by its large volume and complex types,making traditional methods less efficient in classification.Therefore,point cloud automated classification based on deep learning methods has become an important research focus.However,there are still challenges in automated point cloud classification,especially in complex environments.During the acquisition of point cloud data,factors such as occlusion,noise,and sparsity often result in incomplete sample data,which hinders the performance of trained models.Additionally,when dealing with point clouds consisting of billions of points,the difficulty of automated classification increases,leading to lower classification efficiency.In light of these challenges,this paper conducts relevant research and analysis,with the following main contributions:Addressing the lack of large-scale urban scene datasets in current 3D point cloud classification research,this study constructs point cloud datasets of the Nanjing Olympic Sports Center and Suzhou Venture Industrial Park.Raw point cloud data is acquired using a backpack-mounted 3D laser scanner,and undergoes data preprocessing steps such as point cloud registration,segmentation,filtering,and annotation to obtain complete large-scale scene datasets,including training and validation samples.To validate the correctness and reliability of the constructed datasets,three network models based on multi-layer perceptrons,namely Point Net++,Shell Net,and Rand LA-Net,are selected for training and compared with the traditional SVM method.The results demonstrate that on the Suzhou Venture Industrial Park dataset,the classification accuracies for buildings are 98.41%,98.61%,and 92.60%,respectively,while the accuracies for pole-like objects are 73.02%,75.31%,and 74.25%.These findings indicate that the constructed large-scale scene datasets have good point cloud classification performance in three-dimensional urban scenes.To improve the efficiency of automated classification,this paper proposes an RS-Shell Net network model that integrates random sampling techniques into the original Shell Net model,enabling efficient sampling of the raw point cloud data.While ensuring accuracy,this model simplifies the original point cloud data,reducing the amount of data processed by the network model.The results show that on the Suzhou Venture Industrial Park dataset,the classification accuracy for buildings is 93.55%,and the training time is reduced by 24.57% compared to the original model,thereby enhancing the efficiency of network training.