Buildings Extraction From Airborne Laser Scanning Point Clouds Based On Marked Point Process

At the developments of nearly twenty years, airborne laser scanning (ALS) systems have become an accepted means of capturing accurate spatial data in an efficient way because of directly obtain the3D coordinates of objects, short data acquisition and processing times, relatively high accuracy and point density, and reduction in acquisition costs. The ALS data is widely used in numerous applications such as3D city modeling,3D mapping of power lines corridors, digital elevation models generation (DEMs), and environmental studies. Building is an important place for people to live and work. And building is also an important entity in the city space. Data of building outlines are an invaluable data source for automatic generation of cadastral maps,3D building reconstruction, change detection, and so on. Therefore, it is very significant to extract building from ALS data.Extracting buildings from ALS data has been an active topic in the field of photogrammetry, remote sensing, and computer vision. Different approaches for extracting buildings have been reported in the last decade. But due to the diversity and complexity of building and its surrounding terrain, there are still a lot of difficulties in building extraction. So far there is not a good building extraction method which is suitable for any quality data and any scene. On the other hand, due to the discrete randomness and inhomogeneous distribution of ALS data, building extraction is become more difficult. Aiming at these problems, this paper investigates the method of building extraction directly from ALS data based on marked point process. This method is an object-oriented and robust method. Firstly, the Gibbs energy model of building objects is modeled. Then, an optimized solution strategy was put forward to find a maximum a posteriori estimate of the Gibbs energy model. Finally, a refinement operator is performed to extract building roof outlines. The main research context of this paper as follows:1. The research background and significance of this paper as well as ALS system and its application are introduced. After the discussion of characteristics of ALS data and existing problems of building extraction, the research objectives and contents of this paper are proposed. Then the related works of building extraction from ALS data and geometric feature extraction by marked point process are reviewed. Moreover, the difficulties and possible prospects of building extraction from ALS data are pointed out.2. The background and theory of point process and marked point process are introduced. According to the geometric characteristics of buildings, the marked space consists of marked point process of cuboids are defined to describe buildings. Then the Gibbs energy model which can extract building directly from ALS data is build according to the geometric feature and relationship of the objects in the point cloud data. This model contains both a data coherence term which fits the objects to the data and a prior term which incorporates the prior knowledge of the object geometric properties. Spatial relationship and information of buildings are introduced to Gibbs energy model effectively. It laid a foundation for building extraction.3. For the function of Gibbs energy model is a nonlinear function, it is difficult to analysis and obtain analytical solution in theory. Therefore, Reverse Jump Markov Chain Monte Carlo (RJMCMC) coupled with simulated annealing is used to find a maximum a posteriori estimate of the number, the locations and sizes of building objects described by the Gibbs energy model. This solving algorithm can converging to global optimal solution from any initial state. But just as all the sampling-based approaches, the proposed method has also a heavy time cost because of huge computing burden. To improve the efficiency of method, the solving algorithm has been optimized and improved.4. Once the optimal solution of the Gibbs energy U(x) is obtained by RJMCMC coupled with simulated annealing, the building objects of ALS data are detected. However, a lot of objects with the similar shapes of cuboid or with low heights might be detected as buildings as well. On the other hand, one building may have attachments or different levels, resulting in incomplete detection. The above cases lead to bad qualities of building footprints extraction. Hence, a refinement operator is performed to extract building roof outlines from the extracted building objects by filtering false buildings and merging the connecting parts of buildings. Finally, the modified convex hull method is applied to extract the outlines of detected building objects,.5. The proposed model was tested with the ALS dataset of Vaihingen, Germany and Toronto (5datasets in total), in the context of the ’ISPRS Test Project on Urban Classification and3D Building Reconstruction’. This paper analysis and discuss the parameters which were used in the proposed algorithm. And the comprehensive evaluation and comparison on the detected outlines of buildings were performed by ISPRS Workgroup III/4based on the completeness, the correctness, and the quality of the results both on a per-object and on a per-area level. On the other hand, F1measure is used to compare the results of building extraction from the proposed method and other methods. The experimental results show the proposed building extraction method can get good extraction performance in terms of pixel based and object based completeness, correctness, and RMSEs. It also shows that the proposed method provides a functional and effective solution for directly extracting building objects from ALS data of a variety of scenes without resampling or gridding the input data.