Extracting Multi-Dimensional Road Features Towards Safe Vehicle Passage From Complex Point Cloud Data:Framework And Methods

Digitization of as-built road infrastructures is a prerequisite to achieving the life-cycle smart management of road assets and constructing the vehicle-road cooperative system.Light detection and ranging(Li DAR)point clouds(PC),especially mobile Li DAR,can enable accurate and fine depictions of real-world three-dimensional(3D)road environments,which is currently one of the most important data sources that support infrastructure digitalization.However,during the collection of Li DAR data,some nonstationary noises such as pedestrians and cars are inevitably captured.The presence of these non-infrastructure noises may pose significant challenges to the accuracy of identifying and extracting road features.However,a majority of existing studies were conducted in point cloud-modeled road scenes with minor noise issues.Therefore,the adverse impacts of non-stationary noises were not adequately considered.To address this gap,this study focuses on point cloud data with considerable noises.Considering that it is very difficult to extract multiple road features using a single approach,the study aims to develop a framework for automatically extracting key road features towards safe vehicle passage from noisy Li DAR data.First,four crucial road features regarding the passage of vehicles are identified as follows: road edges,vertical clearance,cross slope,and visibility-related information.Based on the fundamental operations that are commonly involved in extracting different road features,a data restructuring approach and a linear encoding-based partitioning method are separately proposed in this study.The data restructure approach can convert complex combined road alignment to a straight line without changing the relative positions of road objects.The partitioning method can segment point clouds into bars,pillars or voxels in a very efficient manner.Besides,the encoding-based partitioning method can establish a binary matrix which helps divide point cloud data into clusters.Through comparisons with mainstream approaches on handing same tasks,the proposed partitioning and clustering method shows remarkable time performance on same computing devices(RAM of 32 GB,CPU of Intel(?) Xeon(?) E5-1650 v4@3.6 GHz).More specifically,when testing on datasets with 4 million to 10 million points,the required partitioning time was reduced from hundreds of seconds to tens of seconds.The proposed method took less than 1s to complete the clustering task while the existing approach may take tens of seconds.Second,based on the fundamental point cloud processing techniques,a similarity and connectivity-based method is proposed to identify road surface points from PC of highways.Test results on two highway sections showed that both the accuracy and recall rate of proposed method exceeded 97%.The road boundaries are delineated and used to filter out non-infrastructure noises.A multiple-coordinate-transformation based approach is developed to fill the missing point regions(MPR)caused by noises.PC of urban streets are partitioned into voxels and the 3D U-net deep neural network is introduced to recognize noises automatically.Through testing on multiple datasets,the trained 3D U-net in the restructured space can accurately segment vehicle,persons and background objects whose Io U measures all exceeded 95%.The road edges in urban scenes are delineated using natural cubic splines.Both road surface and road boundary points are rasterized into binary maps,where image processing techniques are utilized to fill MPR in PC of urban roads.Third,different techniques are separately developed to assess vertical clearance,estimate cross slope,and measure available sight distance on existing road infrastructures,which jointly form the complete framework that reorganizes PC data,remove noises,and extract road features.The effectiveness of the developed framework was demonstrated through tests on multiple datasets.In addition,a procedure for realtime monitoring traffic safety performance(based on the intervisibility)is also proposed in this study which combines static PC of road intersections with dynamic motion data of agents.Through tests on two road junctions,the proposed procedure can finish all the estimations within 0.1 s at each time step,which can support real-time applications.The framework developed in this study is based on LiDAR data with considerable noises.Therefore,it puts less demand on collecting PC during low-traffic-volume hours and can help reduce the long-term mapping cost.The techniques in the developed framework can be applied in extracting different road features from PC of complex road environment.When properly developed,the framework can be applied in the digitalization of as-built road infrastructures.