Multi-Scale Region Growing Point Cloud Filtering Method Based On Surface Fitting

Li DAR(Light Detection & Ranging)is a product of the combination of laser technology and radar technology.It can actively and quickly obtain accurate three-dimensional structural information of the measurement area and form point cloud,which is widely used in the fields of DEM construction,feature extraction and accurate three-dimensional reconstruction.A series of point cloud processing steps are involved.Among these steps,the first step is point cloud filtering.This article mainly studies point cloud filtering,conducts a detailed study on the progressive triangle irregular network(TIN)densification(PTD)algorithm,analyzes its limitations,and makes a series of improvements to the problems:(1)There are gross outliers in the initial point cloud,and the low gross outliers affect the selection of initial ground seed points and then affect the ground fitting.Therefore,a radius filter is used for preprocessing to eliminate gross outliers.(2)A point cloud pyramid is used to organize the point cloud.The sparseness of the point cloud is controlled with different grid sizes.The TIN is densified layer by layer in the iterative process,which optimizes the TIN construction process.(3)In the original algorithm,the initial grid size is large,and the obtained ground seed points are relatively sparse.As a result,the constructed TIN cannot completely cover the point cloud area,resulting in the lack of boundaries in the subsequent judgment process.The method of inserting virtual boundary seed points supplements the boundary of TIN.(4)The PTD algorithm has excessive erosion of the ground during the iteration process.This paper proposes two strategies of "down" growth and "up" growth.After the PTD filtering,the secondary filtering is performed to restore the eroded steep cliffs,mountain bags and other terrain in which the adaptive thresholds are calculated.Ultimately the real ground surface is gradually approached in loop iteration.After the above improvements,by testing the 15 benchmark data sets provided by the ISPRS,Type I error,Type II error,Total error and Cohen's kappa coefficient are 2.40%,3.67%,2.84% and 93.74% respectively,which shows that the proposed method has greater performance to obtain the ideal ground model.