Objects Extraction And Level-of-details Representation From Airborne Laser Scanning Point Clouds

Airborne Light Detection and Ranging(LiDAR)system is comprised of Global Position System(GPS),Inertial Measurement Unit(IMU),Laser Scanner(LS)and others.It could directly capture large-scale,dense,three-dimensional coordinates and physical information from the Earth's surface.Moreover,the laser can penetrate the canopy of vegetation.These merits make it to be a promising tool for capturing terrain and building information.The task of automatically extracting and reconstructing ground and buildings is a prerequisite for many applications.It is the basis of the frameworks of Digital Earth and Smart Earth,and it also is the basis of some professional and non-professional applications,such as urban management and planning,disaster prevention and emergency,environmental simulation,vehicle navigation,tourism publicity.However,airborne LiDAR point clouds have some inevitable characteristics,such as occlusion,data hole,low point density and uneven point density distribution.Moreover,scenes are complex and objects are various in the study area.These characteristics make it challenging to select a suitable element to describe a local surface.Most published methods are not adapted to complex scenes and various terrain types.The preservation of features still remains challenging issues to be settled.Therefore,several novel methods are proposed for robustly extracting ground and building points,and reconstructing high-quality Digital Elevation Models and building LoDs.The main works are described as follows.(1)In general,a single algorithm and a single element are difficult to adapt complex scenes and various terrain types.Therefore,a novel method called a filtering method based on self-adapting partition is proposed.Firstly,the method utilizes pseudo-grids to divide point clouds into grid points and other points.And then,grid points are classified into two classes,which are represented by segment entities and point entities,respectively.Secondly,two classes are filtered by a segment-based filtering method and an improved multi-scale morphological filtering method,and ground grid points are extracted.Finally,a provisional digital elevation model(DEM)is generated from extracted ground grid points,and unclassified points are processed to be ground or non-ground.In order to evaluate the performance of the proposed method,ISPRS benchmark datasets and two large-scale datasets were selected to perform experiments.According to qualitative and quantitative analysis and the comparisons between the proposed method and other excellent methods,the results show that the proposed method is robust for different point densities,diverse terrain types and various complex scenes,and it could preserve terrain features simultaneously.(2)In order to overcome the difficulty that features based on the part of an object couldn't robustly describe differences between various objects,and improve the practicability of algorithms and their adaptive abilities in various complex scenes,the dissert proposes a novel building extraction method based on a top-down strategy.Firstly,object areas similar to buildings are extracted from non-ground points following filtering.Secondly,object areas are classified into two classes,building areas and non-building areas.Finally,the proposed method removes non-building points from building areas.For purpose of evaluating the performance,five ISPRS benchmark datasets and two large-scale datasets were selected to perform experiments.According to qualitative and quantitative analysis and the comparisons between the proposed method and other excellent methods,the results show that the proposed method cloud robustly extract various buildings in complex urban scenes with a good performance in distinguishing buildings from vegetation or other objects,and it could obtain a high precision of building outlines simultaneously.(3)After extracting ground and building points,the dissertation proposes a framework for generating high-quality DEM and building LoDs.In order to improve the quality of DEM,breaklines are extracted by a segmentation-based method and the topological ananlysis of adjacent segments,and building outlines are regularized by a hierarchical method,then breaklines and outlines are taken as constraints in generation of DEM.Moreover,to meet the demands of different LoDs of buildings for various applications,the dissertation proposed a novel method based on morphological scale space.The method creates multi-level point clouds for each building,and a data-driven method is applied to reconstruct the model of each level.Thus,LoDs of each building could be generated.In order to validate the proposed method,two classic datasets were selected for high-quality DEM generation and building LoDs reconstruction.According to qualitative and quantitative analysis,the results show that the high-quality DEM with constraint of terrain features could be generated.The proposed could directly generate the model of each level from the raw building point clouds.The coarser model does not need to be derived from the finer model by simplification and generalization,thereby the proposed method could better serve the demands of different users.