| Road aggregates are closely related to road transportation,and their performance and quality directly affect the safety and durability of roads.Therefore,it is of vital importance to analyze and evaluate aggregate data.This thesis proposes an optimized point cloud registration method based on PCRNet,aiming to improve the accuracy and efficiency of aggregate 3D point cloud data registration,provide reliable data support for aggregate gradation prediction field,and thus improve the performance of pavement,and help develop green preventive maintenance.Firstly,this thesis uses high-precision 3D laser scanning equipment to obtain the depth images of aggregates,and improves the data quality,and then completes the construction of standardized point cloud data set.To address the noise problem during the depth image acquisition,the ASF alternating sequence filtering algorithm is used to denoise the depth images.Then the depth images are processed by threshold segmentation and format conversion techniques to obtain the 3D point cloud data of a single aggregate.The point cloud is downsampled using the farthest point downsampling algorithm to reduce the amount of point cloud data while retaining key feature information.The final 3D point cloud dataset of 4400 aggregates is obtained.Secondly,this thesis improves the basic point cloud alignment network PCRNet.To address the problems that PCRNet has difficulty in extracting multi-scale features and low efficiency in obtaining rigid body transformations using fully connected layers,this thesis introduces Position Adaptive Convolution(PAConv)for feature extraction and replaces fully connected layers with Singular Value Decomposition(SVD)module to build the PACNet network based on the above improvements.And the performance of the improved module is verified by ablation and comparison experiments.The experimental results show that the error metrics(MSE,RMSE,MAE)of both rotation matrix and translation vector are reduced by more than 26.48% compared with the original PCRNet,and the error metrics of rotation matrix are reduced by more than 5.77% compared with FGR,which has lower error among the mainstream alignment methods(ICP,Go-ICP,FGR,PointNet LK,DCP).Finally,the PACNet network is further improved.In this thesis,the improved Transformer module is added on the basis of PACNet network,which can capture the correlation information in the point cloud and perform feature fusion on the extracted features,based on which the Trans-PACNet network is constructed.The ablation experimental results show that all error metrics are reduced compared with the PACNet network and the original PCRNet network.The comparison experimental results show that the Trans-PACNet network proposed in this paper reduces all error metrics by more than 20.54% compared to the lower error FGR for noiseless,Gaussian noise and randomly cropped data;compared to the Go-ICP method with the fastest alignment speed among the non-learning methods,the alignment of 512,1024,2048 Compared with the Go-ICP method,which is the fastest non-learning method,the time required to align point clouds of 512,1024 and 2048 resolutions is reduced by 254,417 and 810 milliseconds,respectively.In summary,based on the 3D point cloud data of roadway aggregates,the proposed point cloud alignment method can significantly improve the alignment quality,possess stronger robustness and generalization,and achieve faster alignment speed compared with the mainstream alignment methods. |
