Mobile Laser Scanning Point Cloud Classification Based on Data Augmentation and Mask Learning
Objective A mobile laser scanning(MLS)system can quickly and accurately acquire 3D point cloud data of a scene around a road,and such a system has the advantages of fast acquisition speed,high positioning accuracy,high point cloud density,strong anti-interference ability,and information richness.Thus,MLS systems are widely used in many fields,such as intelligent transportation,digital twin cities,high-precision maps,and assisted driving.As the basis of this system application,people need to extract accurate and semantically rich information from large-scale and complex MLS point clouds.Although there are many large-scale deep learning point cloud classification methods that achieve competitive classification accuracy,they still suffer from problems such as insufficient original training scenes and incomplete point cloud feature expression.To further expand the distribution of the training data and improve the feature representation of the model,we propose an MLS point cloud classification method based on data augmentation and mask learning.Methods The proposed method is divided into two main parts:an elevation-calibrated Mix3D(EC-Mix3D)data augmentation strategy and a mask learning framework.Specifically,the EC-Mix3D data augmentation strategy first extracts the normalized elevation of the point cloud through a cloth simulation filtering algorithm.Then,the point clouds of two independent sub-point clouds are mixed via normalized elevation calibration.Because the point clouds of the two scenes have the same elevation reference,they can be effectively mixed to generate training data with a new context,which expands the distribution of the training data.The mask learning framework first generates mask data by applying a random block mask operation to the input data.Then,the original input data and mask data are fed into the Siamese two-branch network and predicted.Finally,label supervision,consistency constraints,and error prediction entropy maximization based on two-branch predictions are performed to enhance the expressiveness of the point cloud features and reduce the model's overconfidence when predicting in complex regions.Specifically,the label supervision of the original branch provides the underlying supervised signal for model training,and the label supervision of the mask branch can motivate the model to efficiently learn additional point cloud context features by predicting masked data from unmasked data.Consistency constraints increase point cloud feature expressiveness and model prediction stability by minimizing the difference between the predicted values of the masked input branch and the original point cloud branch,and error prediction entropy maximization increases prediction uncertainty in complex scenarios by encouraging high entropy posterior probabilities for misclassified points,which suppresses classification model overfitting.Results and Discussions The Toronto3D and Paris public MLS datasets were used to validate the proposed method.Experimental results(Fig.6,Fig.7)show that the proposed method can effectively classify most MLS point clouds.However,the classification results for points far from the center and feature similarities are unconvincing.Comparing these results(Table 1,Table 2)with those obtained via other methods shows that the proposed method can effectively improve the accuracy of the baseline method and obtain optimal accuracy.Specifically,we obtain 97.7％and 93.20％overall accuracy(OA),and 83.8％and 68.74％mean intersection of union(mIoU)on the Toronto3D and Paris datasets,respectively.The ablation experiment results(Table 3)show that each component of the proposed method improves the classification performance of the model.By sequentially adding the EC-Mix3D data augmentation strategy and mask learning framework to the baseline method,the mIoU can be improved by 3.38 percentage points and 6.34 percentage points,respectively.The results of different mask voxel sizes(Table 4)and mask ratios(Table 5)indicate that the proposed method employs the mask parameters that are the most suitable for Toronto3D point cloud classification.Finally,the results of a model complexity analysis(Fig.8)show that the proposed method enhances model classification performance without increasing model complexity.Conclusions In the present study,a novel MLS point cloud classification method is proposed to improve model robustness and the expressiveness of the point cloud features via an EC-Mix3D data augmentation strategy and a mask learning framework.Specifically,the EC-Mix3D strategy expands the training sample distribution by scene mixing the training data to improve the robustness of the model.The mask learning framework obtains a masked point cloud by applying a random block mask to the input point cloud,and it then utilizes the predicted values of the original and masked point cloud for label-supervised learning,consistency constraints,and error prediction entropy regularization,aiming to enhance the feature representation capability of the model.Two public MLS datasets(namely,the Toronto3D and Paris datasets)are used to validate the proposed method.The results show that the proposed method can effectively classify MLS point clouds and obtain mIoU of 83.8％and 68.74％,respectively,which are better than those of other methods.
vehicle-mounted mobile laser scanning point cloud classificationEC-Mix3D data augmentationmask learning frameworkconsistency constrainterror prediction entropy maximization
万方数据
维普
