基于激光雷达点云的道路几何信息提取与数字化建模研究

Extraction and Digital Modeling of Road Geometric Information Using LiDAR Data Point Clouds

为实现快速、自动化的道路几何信息提取和数字化建模,基于激光雷达点云提出了一套从道路语义分割、几何线形提取到集成化建模的通用框架。首先,基于空间上下文特征基础框架,将局部特征的最大值和邻域均值进行聚合以作为局部特征,使用径向分布参数与三维坐标描述全局上下文特征,构建道路语义分割网络。其次,基于道路场景分割结果,通过体素降采样和半径滤波法减少点云数据量、去除离群点,并利用可变半径Alpha Shapes(VA-Shapes)算法提取道路边线,结合获取的边线横纵坐标,计算路段几何信息(路宽、纵坡、横坡等),使用inshape函数和插值法构建交叉口的数字高程模型。最后,采用Dynamo for Revit将道路几何信息导入并生成道路路线,通过Revit软件设计道路自适应族构件及不同类别基础设施族构件,实现精细化道路数字建模。利用开源数据集Semantic3D进行训练和测试,分析与评价道路几何信息提取效果。研究结果表明:所提出的算法总体准确度为95%,路面的单类交并比为97.9%,能够很好地实现道路点云场景的自动化语义分割;相比于传统的固定半径Alpha Shapes算法,VA-Shapes算法的时间复杂性较低,可以较好地提取道路边线;通过提取算法与手动测量结果的对比,显示不同几何信息的平均绝对误差较小,说明算法具有可靠的精度。提出的“点云语义分割-几何信息提取-BIM数字建模”过程实现了逆向构建数字化道路模型,对现役道路基础设施智能管理具有重要意义。

A general framework from semantic segmentation to geometric information extraction and integrated modeling was proposed based on LiDAR data to rapidly and automatically extract road geometric information and complete digital modeling.The local maximum and neighboring point mean features were concatenated as local features based on a fundamental foundation of spatial contextual features.Three-dimensional coordinates and radial distribution were combined to describe the global contextual features,and a semantic segmentation network was established.Additionally,the voxel grid filter and radius outlier removal methods were used to minimize the amount of point cloud data and remove outliers.The adaptive radius variable alpha-shapes method(VA-Shapes)was then employed to extract the road boundary based on semantic segmentation results.Furthermore,the geometric data of the road,including the road width,longitudinal gradient,and cross-slope,were obtained from the horizontal and vertical coordinates of the boundary.The in shape function and interpolation method were then applied to establish a digital elevation model.Subsequently,road routes were generated from the extracted road geometric information using Dynamo for Revit,and adaptive road components and various infrastructure components were constructed using Revit,developing a detailed digital road model.The Semantic3D dataset was utilized for training and testing to analyze and evaluate the extracted road geometric information.The overall accuracy(OA)of the proposed net is 95%,whereas the intersection-over-union(IOU)of segmented pavement is 97.9%,indicating that the proposed net could accomplish superior performance on semantic segmentation of point clouds.Compared with the traditional fixed radius A-Shapes method,the temporal complexity of the VA-Shapes method is low.In addition,the VA-Shapes method can efficiently extract the road boundary.The mean absolute errors between the extracted and manually measured geometric information are slight,demonstrating the effectiveness and accuracy of the proposed methods.The proposed process from semantic segmentation of the point cloud for geometric information extraction and building information modeling for digital modeling has the potential to build a digital model of a road in reverse,which is critical for the intelligent management of existing road infrastructures.