MODELING OF ROOFS FROM POINT CLOUDS USING GENETIC ALGORITHMS

Abstract: Building roof extraction has been studied for more than thirty years and it generates models that provide important information for many applications, especially urban planning. The present work aimed to model roofs only from point clouds using genetic algorithms (GAs) to develop a more automatized and efficient method. For this, firstly, an algorithm for edge detection was developed. Experiments were performed with simulated and real point clouds, obtained by LIDAR. In the experiments with simulated point clouds, three types of point clouds with different complexities were created, and the effects of noise and scan line spacing on the results were evaluated. For the experiments with real point clouds, five roofs were chosen as examples, each with a different characteristic. GAs were used to select, among the points identified during edge detection, the so-called ‘significant points’, those which are essential to the accurate reconstruction of the roof model. These points were then used to generate the models, which were assessed qualitatively and quantitatively. Such evaluations showed that the use of GAs proved to be efficient for the modeling of roofs, as the model geometry was satisfactory, the error was within an acceptable range, and the computational effort was clearly reduced.

摘要：建筑屋顶提取的相关研究已开展三十余年，其生成的模型可为诸多应用场景（尤其是城市规划）提供关键支撑信息。本研究旨在仅利用点云（point cloud）数据，结合遗传算法（Genetic Algorithms, GAs）开发一种更自动化、高效的屋顶建模方法。
为此，本研究首先开发了一种边缘检测算法。本研究针对激光雷达（LIDAR）采集的模拟点云与真实点云开展实验：在模拟点云实验环节，构建了三种不同复杂度的点云样本，并评估了噪声与扫描线间距对实验结果的影响；在真实点云实验环节，选取了五栋具备不同特征的屋顶作为测试示例。
本研究借助遗传算法，在边缘检测阶段所识别的点集中筛选出所谓的"显著点"——此类点是精准重建屋顶模型的核心要素。随后利用这些显著点生成屋顶模型，并分别从定性与定量两个维度开展模型评估。
评估结果显示，采用遗传算法开展屋顶建模具有良好的有效性：模型几何形态符合预期，误差处于可接受范围内，且计算量显著降低。