Filtering By Transfer Learning And Building Feature Lines Extraction Based On Airborne LiDAR Point Cloud Data

Airborne LiDAR(Light Detection And Ranging)has been becoming an indispensable and active remote sensing technique concerning three-dimensional geospatial data acquisition of the Earth's surface.The high-density and high-precision point clouds data acquired by airborne LiDAR has also been becoming one of the important source data for object identification.Filtering,a process of ground point's extraction,is one of the key steps in airborne LiDAR data process.It is also the prerequisite and foundation of other objects' detection.Traditional filtering methods always need to preset different parameters on the basis of some common knowledge and experience,according to different terrain topography.This may result in limited expansibility of these filtering methods.In recent years,many researchers developed some filtering algorithms combined with a certain classifier and obtained some reliable results.Compared with the traditional high-resolution images,LiDAR data has less directly available information for object detection.As a result,feature generation is a necessary step in this type of filtering.However,increasing number of features generated from geometric information also brings in some new problems.Unfortunately,more features do not necessarily guarantee higher classification accuracy.On the contrary,too many features not only bring a lot of irrelevant and redundant information to a classifier,which in many cases decrease the classification accuracy,but also increase computational cost.Thus,feature selection is urgently needed.Moreover,the current filtering methods set parameters or select samples only for the new dataset,without taking old classified datasets checked by experts into consideration.How to explore the knowledge from the old datasets to help classifying the new one effectively and reduce unnecessary labor cost is a meaningful research.In addition,buildings are also one of the most important objects in airborne LiDAR object detection.The existed research work focus on building detection and three-dimensional modeling.But there is almost no research related to feature lines on building roofs,such as ridge lines and valley lines.And due to complex scenes and various building structure,many related algorithms cannot obtain high accuracy and degree of automation.Consequently,the aim of this thesis is to propose a building process,which includes a building detection method and a feature lines extraction algorithm.In this thesis,we deeply investigate some critical algorithms related to filtering and building detection using airborne LiDAR data.The major works are listed as below.(1)It is lack of researches about feature selection for airborne LiDAR data filtering.A filtered feature selection algorithm based on Parzen window estimation is adopted in this thesis.The objective is to select the most relevant features for filtering urban or rural scenes and to provide an accurate filtering with a small number of features.Parzen window estimation has also been introduced herein to optimize the estimation of joint probability density between continuous values and discrete values.(2)The traditional filtering methods need to preset parameters varied with different terrain topography.The thesis focuses the emphasis on filtering with a certain classifier and compares three supervised classifiers for point cloud filtering.The features are generated 'from the geometric information of point clouds and selected by the proposed feature selection method before.In addition,tentative exploration in the application of transfer learning theory is also made to point cloud filtering.Considering the terrain and data difference between source dataset and target dataset,active learning mechanism is also introduced in this thesis to reduce labelling cost and improve filtering accuracy of target dataset.(3)Extracting feature lines from the raw airborne LiDAR data is a complicate process.A rapid feature lines extraction workflow is proposed,which is on the basis of the point clouds.Firstly,a grid-based method is adopted to extract edge points.According to building geometry,the positions of corner points can be determined to categorize edge points into a certain number of groups.Then building outlines can be fitted by each point group and refined through building geometric knowledge.Secondly,a climbing algorithm is proposed to search horizontal ridge points.Then the relationship between points is analyzed to determine the endpoints of slant ridge lines and valley lines.