Building Area Extraction From Airborne LiDAR Point Clouds Using Mathematical Morphology

Building area extraction has long been a hot research field of photogrammetry, many scholars have tried to extract building areas just from the images at the very beginning, but it is very difficult to fulfill it automatically due to the complex texture information of buildings. Generally, human intervention is required, especially for these large scale imagery. More and more scholars are now concerned about Airborne LIDAR technology since it becomes hard to extract building regions only from aerial images. Using airborne LiDAR technology can obtain the 3D information of buildings so as to change the task of building extraction process into geometry information, which would contribute to the automation of building extraction method. In recent years, along with the development of dense matching algorithm, intensive and high accurate 3D point cloud data can be generated from stereo aerial imagery, but compared with the three-dimensional laser point clouds, dense matching point cloud would produce blank area in weak texture region, while point clouds in strong texture region are too much thick, which would cause the imbalance of dense matching point cloud to increase the difficulty in handling the point clouds. Besides, dense matching point cloud cannot give echo information.Mathematical morphology is a spatial structure analysis theory which is based on some mathematical theories about ensemble, integral geometry and algebra. This theory is mainly used to analyze the shape of an object and its composition; it can also be used in the process of 3D point cloud data processing, since the processing of 3D point cloud is the processing of geometric information of objects. At present, many scholars take advantage of this theory to filter the 3D point cloud and to generate 3D point cloud ground and non-ground information. This paper mainly talks about the extraction of building area of airborne laser point cloud based on the mathematical morphology, trying to achieve automatic and accurate extraction results of the building area. The main contents are as follows:1) In order to apply mathematical morphology in laser point cloud data, the laser point cloud need to be interpolated to generate the grid data. During the process of interpolation, grid spacing is consistent with that of point cloud and the quantity and number of point cloud in each grid should be recorded, which would keep the original information of laser point cloud. Top-hat transform in mathematical morphology is brought about to remove high and low noise points, during this process, whiter top-hat transform is used to remove high noise points while black top-hat the low noise points. Apart from this, in order to facilitate the differentiation between the buildings and the vegetation areas, the vegetation area of the study area was identified by using the vegetation index of imagery and the multiple echo attributes of laser point cloud.2) A method based on watershed algorithm for the extraction of Airborne LIDAR point cloud building region is proposed. First, generating the morphological gradient of laser point cloud grid data; then, sorting the gradient data using the bucket sort algorithm; finally, immersing the laser point cloud data according to the gradient sort order to complete the segmentation of the laser point cloud. In the process of segmentation, in order to separate the vegetation area connected to the building, the immersion process is restricted by the vegetation zone mask. After segmenting the laser point cloud, the segmentation region is classified by the evidence theory, and the data source is mainly generated by four kinds of data sources:(1) the difference between the elevation of the current region and its adjacent area, (2) the percentage of the adjacent area whose height is less than the average elevation, (3) the vegetation index, (4) the proportion of the points of the multiple echoes. Using fuzzy theory to describe the probability of four kinds of data sources, and can find the joint probability of these four by evidence theory. According to the classified results of laser point cloud, the area threshold is set to extract the building area.3) Proposing an airborne laser point cloud building extraction method based on connected operators, and combining it with that method based on watershed algorithm. First, threshold superposition is performed for the laser point cloud data, the binary image is generated at different elevation levels; then, making contraction-based second-generation connectivity analysis for the binary image in different elevation level to elimate the "weak link", and taking advantage of the mask-based second-generation connectivity to remove the vegetation area connected with the building; finally, the ground area is filtered out and building area is identified by using the difference of area deviation between the building and the ground in the adjacent elevation level. In the process of extracting buildings by connected operators, for some buildings with low elevation and large slope, it is difficult to be separated from the ground in the process of connectivity analysis, so that it cannot be identified. However, for some dense and rough buildings, it is difficult to be extracted using the method based on watershed algorithm, while can be achieved by the extraction method based on connected operators. The analysis above shows that, watershed algorithm and connected operators have their own respective advantages for different types of building extraction, thus, combing these two methods is considered in this paper.4) Making accurate evaluation and analysis on building extraction results. The accuracy evaluation of the building extraction results by watershed algorithm and connected operators and the final building region are performed based on the pixel and object accuracy evaluation method, and the influence of building areaTand overlap threshold on the accuracy evaluation results is analyzed. What's more, the accuracy evaluation of the results of building extraction from five text areas were carried out, and the results were compared with that of other methods in ISPRS website.