Research On Urban Road Extraction Algorithm Based On Airborne LiDAR

Roads are the "vessels" and "skeletons" of the city,connecting and supporting the development of the city.Roads are also closely related to people's daily lives.It is one of the components of GIS data.Road information is also an important data source for "smart traffic".Accurate road extraction is very important for a city.Airborne LiDAR has become one of the important data sources for urban road extraction due to its characteristics of being not affected by weather conditions,fast operating cycle,strong timeliness,high degree of automation,and certain vegetation penetration.Airborne LiDAR urban road extraction involves a series of steps,among which point cloud filtering and road contour and centerline extraction are particularly critical.Point cloud filtering is the process of separating ground points and non-grounds.The traditional filtering algorithm mainly sets filtering conditions based on the spatial position relationship.Although it is more mature,the overall universality is still insufficient.With the development of computer vision and machine learning,machine learning can find more useful information for point cloud filtering.The application of machine learning to point cloud data processing is also a hotspot for future research.But the general network is time-consuming and the data also needs complicated preprocessing.Therefore,how to conveniently and quickly train the network to achieve the effect of point cloud filtering is a hotspot of research.For road extraction,remote sensing image extraction is susceptible to occlusion of buildings and vegetation,and the point cloud data is more fuzzy than road boundaries.How to combine the characteristics of the two and complement each other to achieve better road extraction has become a research hot spots.This paper has conducted in-depth discussion and research on cloud filtering of important ring nodes for road extraction and road contour and centerline extraction.The main research contents and research results are as follows:1.In this paper,the idea of moving surface fitting filtering is applied to point cloud denoising.Setting a threshold can remove terrain points while removing noise points,but due to the influence of window size,local surface fitting is easy to misclassify cluster noise Point,and the operation is more cumbersome.In this paper,an improved surface fitting denoising algorithm is proposed.First,the grid is divided into point clouds and the seed points are determined.The partial least squares surface fitting parameter estimation is performed by using the model with some variables and errors to calculate the vertical distance from each point to the surface Use the frequency distribution histogram to eliminate noise points.Two sets of data with high and low noise points are selected to experiment with the proposed algorithm.The experimental results show that the noise points can be removed efficiently and quickly,which can meet the requirements of subsequent point cloud data processing.2.Use PointNet ++ to filter the airborne LiDAR point cloud.The grid points have been calibrated for the location data,and the grayscale information is added for training.The training classifier can get the corresponding network parameters.The test data can be filtered.Using 15 sets of test samples to test the filtering effect and a set of data containing color information issued by the International Society of Photogrammetry and Remote Sensing,the algorithm was experimentally analyzed to verify the filtering accuracy of the proposed method in various complex environments.The results show that most of the existing mainstream filtering algorithms of this algorithm are similar,in which the average Kappa coefficient is 78.31%,and the filtering effect of individual flat areas is better than that of the mainstream algorithm.It removes the ground points while retaining the details of the ground.Incorporating color information can improve the filtering accuracy.3.Proposed a road extraction method of Gaussian mixture model combined classification.This method uses fusion images,that is,point cloud data containing color information.Firstly,the gray information and properties of point density in point cloud data are classified by the Gaussian mixture model for pattern classification to extracted the road seed region,which is expanded and constrained by the intensity image.Finally,the Snakes and Mathematical morphology are used to extract road centerline efficiently.To verify the effectiveness of the method,two sets of LiDAR point cloud data located in a foreign city were used for experiments.The results show that this method can effectively reduce the impact of shadow occlusion on road extraction,the centerline of the extracted road is relatively smooth,the overall integrity rate of the extracted road is 93.96%,the average accuracy rate is 93.57%,and the average quality is 88.30%.