Filtering Method Of Airborne LiDAR Data Based On Category Property Of Point Clouds And Terrain Structure Feature

With goble positioning system(GPS) and inertial navigation system(INS) combined to improve positioning accuracy, generating digital terrain models(DTMs) from airborne LiDAR data had become very mor and more popular. And compared to traditional photogrammetry, data collection of airborne LiDAR was less requirements on the weather, season, and time requirements of small and quick achieve to the suface three-dimensional spatial information. The key issue of constraints airborne LiDAR data processing for a long time was automatic classification of discrete point clouds. Point clouds classification for actual application could be divided into point clouds filtering, that was, non-ground point and ground point, as well as buildings, vegetation and other surface features information extraction and so on. In the last 20 years, automatic classification of point clouds was a front and problems question in the fileds of LiDAR data processing and photogrammetric data processing.In recent years, the theoretical methods of airborne LiDAR data processing were developing rapidly. Some of these method had been widely circulated into practical products. For example, TerraScan which developed in TerraSolid company of Finland was considered as a representative of the production of airborne LiDAR point cloud software with the most widely used. SCOP++ of Inpho company in Germany has better performance in dealing with discrete point clouds from the traditional photographic measuremen. Extraction information of surface features, the German definiens company's eCognition with object-oriented information extraction from raster image was undoubtedly most people encouraging results in the last 10 years. In this paper, according to depth research with most traditional algorithms, some improvement or innovative approaches have been proposed in point clouds preprocess,filtering and classification and so on. The key problems and technical details of these algorithms in implementation and practical analysis have been discussed detailedly. These elements include:1) Noise removing method of point clouds based on 3D finite element analysis was proposed. With the analysis of limitations of some traditionnal noise removing algorithms, such as local adjacent points fitting, frequency domain signal analysis, and combined with the spatial distribution of the typical noise in LiDAR point clouds, a noise removing algorithm based on finite element analysis has been proposed. Firstly, it used space hexahedral model as the basic "finite element", choosed a certain threshold for rapid subdivision of the original LiDAR point clouds to three-dime-nsional mesh, and constructed the topology relations for discrete point clouds. Secondly, noise units and non-noise units were clustered with reasoning rules via adjacency relations, and noise units and non-noise units were distinguished effectively.Finally,residual noise points were removed by iteration calculation by using of more fine threshold value for subdivision to achive better result of noise removing. Based on comparison of finite element analysis method and local adjacent points fitting method with multi-groups experiments from typical LiDAR point clouds, it was shown that finite element method has obvious advantage in removing typical noise.2) Point clouds rough classification based on Information extraction technology in raster images was proposed. An automatic classification method for urban LiDAR point clouds was proposed. Firstly, height texture image was segmentated with a over-segmentation resistance segmentation mehtod named topology heuristic segmentation algorithm. Secondly, ground polygon objects and non-ground polygon objects were divided by iterative Otsu clustering theory. And then some adjacent non-ground objects were merged naturally. Finally, typical objects were detected by using of two topology models. And tree objects were detected with the new feature of multi-echoes from typical objects.What's more, building objects and other objects were distinguished with additional conditions of region area and height of buildings. Multiple experiments of city point clouds data with a mass of feature types were done in the last term. And it was shown that:the proposed algorithm has good classification accuracy and strong practical value.3) Filtering method of urban LiDAR point clouds under the guide the class properties of point clouds was proposed. Considering to the problems of urban LiDAR point clouds, a filtering algorithm with adaptive TIN models under the guide the class properties of point clouds was proposed. Firstly, the raster image from grid interpolation was segmented by using of object-oriented segmentation. Secondly, ground polygon objects and non-ground polygon objects are divided by iterative Otsu clustering theory and topology models. Finally, the initial terrain triangular mesh was constructed from classification results. And then, to achieve the high quality filtering results, the parameters criterion of terrain points were adjusted adativly. Multiple experiments of filtering for real city point clouds data were done in the last term. What's more, accuracy evaluation with traditional method shows that the proposed algorithm has better performance for fitering of urban point clouds.4) Fitering method of mountainous areas point clouds with taking into account structural features was proposed. For purpose of persisting topography features in the filtering of mountainous areas, a filtering algorithm with adaptive TIN models based on analysis of contour shape features was proposed. Firstly, contour lines with regular interval were tacked in original terrain data, and some smoothing process were done. Secondly, the split algorithm could calculate sharp feature points including ridge poings or valley points one by one. And the total terrain region can be divided by a quadtree. Finally, the initial terrain triangular mesh was constructed with seed points from the rectangle regions of quddtree nodes. And terrain triangular mesh was established increasingly in the next step.Multiple experiments of filtering for different systems point clouds data were done in the last term. What's more, accuracy evaluation with traditional method shows that the proposed algorithm could improve filtering quality of mountainous areas.All of the point clouds data processing theory, methods and algorithms proposed in this paper were practiced with C language in windows environment. Experiments were done by selecting a lots of groups data from different LiDAR systems. And the results were shown that the point clouds processing methods proposed in this paper were good use to various airborne LiDAR system. And it was can be fit to many real scene data sets with different types of manmade objects and nature objects. Additionly, the rough classification method of point clouds based on class properties of point clouds had good performance in extraction of buildings and trees. The filtering method discussed in the last term could be better than the traditional correlation method, and could meet the actual mapping requirements better.In this approach, noise removing method, rough classification method and filtering method considering class properties of point clouds and topographical features were very practical approaches. In the theoretical sense, they were a set of new method which could deal with original LiDAR point clouds effectively, and could also be provided with forward looking and sustainability, such as developing modeling methodologies combining with spectral imageing, and researching methods of true-orthomap production. In the practical sense, some algorithms proposed in this paper could be applied to airborne LiDAR system or digital photogrammetry system directly.