| The principal premise of automatic 3D building reconstruction is to detect the building's location and extract its roof boundary precisely. The result of detected buildings also have a high practical value, which can be applied in fields such as urban planning, land investigation, change detection, military reconnaissance and solar energy potential evaluation. Since the airborne platform tends to obtain large amounts of information of the roof, researches on building detection based on airborne data have been given much attention by domestic and international academics over the past few decades. Among the data sources used in building detection, aerial images and LiDAR (Light Detection And Ranging) are the most representative data. The automatic detection and precisely contour extraction of building have always been a complicated research subject. The difficulties of this kind of researches mainly based on LiDAR data, for instance, can be summarized as follows: ?The terrain elevation information obtained through LiDAR filtering is one of the most important bases to separate buildings from other ground objects, but although the existing filtering algorithms are albe to meet the application requirements of building detection to some extent, the problems including low degree of automation (especialling when processing data with persistent noise) and poor performance on processing particular complex terrain; ?Roof structure of the building in the real world is becoming more and more complicated, better strategies are required to make reasonable assumptions, and to extract correct building contours when facing the interference of other groung objects, especially the trees locating around the builidngs; ?The quality of building contour extracted from LiDAR data is relatively low because of the limitation of scan resolution. Although arieal images contain a wealth of spectral and edge information, using images directly for building detection and contour extraction will not only be influenced by the shadow, occlusion and the lack of texture, but also have to face the problem of low degree of automation. Therefore, a suitable solution has to be proposed to combine the advantages of the two data sources, which aims to enhance the automation degree of building detection as well as to improve the quality of the extracted building contours.Aiming at the above key issues, researches have been carried out from three aspects in this paper:LiDAR filtering, automatic building detection based on LiDAR data, and building boundary refinement based on aerial images. Therefore, a relatively complete and feasible technical solution of precise extraction of building boundary can be obtained. The main work of this article can be concluded as follows:1) Research on point cloud filtering method based on modified progressive TIN densification with many improvements. Classical approaches tend to obtain seed points by selecting the lowest point in a local range. The noise low point will be selected as the seed point inevitably without pretreatment, thus influencing the final filtering result greatly. A conventional way to remove the noise low point is to analyze the height histogram and isolated point. However, points close to the ground and appear gathered are still difficult to remove without human interference. In order to solve this problem, two improvements were made in this paper. On the one hand, the theory of confidence interval estimation were introduced to analyze the seed points obtained from the initial selection, and seed points that do not meet the terrain assumption were eliminated. Meanwhile, new reliable seed points in the vacant area were selected iteratively to assure that all seed points meet the check condition. Therefore, optimization selection of seed points can be realized without preprocessing. On the other hand, a rule-constraint region growing approach was proposed to detect sharp ridges. The rule, specifically, is that if a triangle is suited to the feature descriptions of sharp ridge topography during TIN densification, region growing was implemented on the current triangle so that the complete topography can be detected; for slope terrain fully covered by ground objects, a two-stage filtering technique was adopted to get both correct and complete results. The initial filtering was performed with a relatively loose threshold, and then the second one was preformed for the ground points obtained with a conservative threshold. Experimental results indicated that the proposed approach had a strong ability of filtering, in the statistics of evaluation indicators on 10 representive test regions, the proposed method had a total average error with 3.71%, and an average Kappa coefficient with 91.13%, which exceeded the other six comparative approaches.2) Research on building detection method from LiDAR data based on plane feature constraint. In the real world, most structures of roof building can be split into several planes, which is the basic assumption of many existing detection approaches. In this paper, the automatic building detection method based on plane feature constraint were implemented and improved. The proposed approach needs neither auxiliary of multispectral images nor rasterization of LiDAR data, but only takes original point cloud as the processing data. First, KD tree structure of discrete point cloud were constructed according to their coordinates, and then plane fitting based on singular value decomposition was performed to extract 3D planar patches. The patches close to the ground were eliminated by the elevation information provided by digital terrain model. Meanwhile, the main plane structure of the roof can be obtained by merge operation for the extracted patches. The following step was to conduct region growing based on plane structure, taking the other points, which is in the neighborhood of the main plane, as judgment objects. Finally, the planes were clustered by a certain constraint, and Alpha Shapes algorithm was adopted to extract the boundaries of the clustered building point cloud. Results of 3 open datasets from international society of photogrammetry and remote sensing manifested that the proposed approach were able to extract contours for many kind of building roofs. In the comparison of key indicators, which is the average quality of 3 datasets, the proposed method reached 88.31%, outperforming all the other comparative methods.3) Research on building boundary refinement method based on modified Snake model. Since the precision of building boundary obtained from LiDAR data is restricted by its relatively low scanning resolution, the precision can be improved with the help of the texture information, especially edge information from high-resolution images. Different from common approaches which use the edge line segments of the roof extracted from images to replace and update the original boundary, the proposed refinement method avoided the procedure of line feature extraction by using modified Snake model. Refinement strategy in object space using multi-view images was adopted to refine initial boundary obtained from LiDAR. Reasonable refined boundary can be obtained by firstly developing energy function suitable for building boundary with the constraints of the deviation angle, image gradient, and area, and then actively moving the nodes of the boundary in a certain range to find the best optimized result using greedy algorithm. Considering both precision and efficiency, the candidate shift positions of the boundary nodes were constrained and the searching method was optimized. The experiments showed that the proposed strategy of building contour refinement was effective and feasible. The quality index of the building contour was improved in both 3 datasets, while the key evaluation index, which was average quality index, was improved from 91.66% to 93.34%. The statistics of the evaluation results for every single building demonstrated that 77.0% of the total number of contours were updated with higher quality index. |
PDF Full Text Request