Lidar Nokta Bulutu Verisi ve Yüksek Çözünürlüklü Ortofotolar Kullanarak Bina Çıkarımı İçin Bir Yaklaşım

Abstract

Nowadays, with the development of sensor technologies in remote sensing, there has been a significant increase in object detection studies. Particulary, building detection is among the common fields and important applications from remote sensing data such as LiDAR (Light Detection and Ranging) point cloud data and high spatial resolution orthophotos. In this study, a method based on Hough transform, perceptual grouping and seeded region growing has been developed for automatic building extraction and reconstruction from high resolution color (Red, Green, Blue) orthophotos and LiDAR point cloud data. The first stage of the method is pre-processing, which includes the registration of LiDAR data and orthophotos, noise removal and ground filtering from LiDAR point cloud data. Then, Digital Surface Model (DSM), Digital Terrain Model (DTM) and normalized Digital Surface Model (nDSM) are generated from LiDAR point cloud data, and VARI (Visible Atmospherically Resistant Index) vegetation index is generated from orthophoto. A threshold is applied to nDSM in order to separate the high vegetation areas and buildings from the low height objects. Next, the vegetation areas are masked from the thresholded nDSM by using the vegetation index band and therefore only the building areas remained. After detecting the building areas, the edges are extracted from orthophoto with the DoG (Difference of Gaussian) filter. Line segments that form buildings are extracted from the obtained edge image using Hough transform, and the building boundaries are constructed from these line segments using the developed perceptual grouping rules. The areas within the constructed building boundaries are then taken as the seed regions and buildings are detected using the seeded region growing segmentation operation. The method was applied on ten test fields with different characteristics selected from the city of Bergama, Turkey. Accuracy assesments of the obtained results were carried out by comparing them with the reference data which was generated by manuel drawing from orthophoto. Based on the obtained results of the perceptual grouping algorithm, the pixel-based and object-based average accuracy values were%82.56 and %96.75 for BDCom. (Building Detection Completess) and %84 and %100 for BDCor. (Building Detection Correctness), respectively. Based on the results of pixel-based accuracy analysis, the average BDCom. and BDCor. accuracy rates of the combined perceptual grouping and seeded region growing segmentation method were %89.82 and %96.37, respectively. Based on the results of object-based accuracy analysis, the average BDCom. and BDCor. accuracy rates were %84 and %100, respectively. According to the results of pixel-based accuracy assessment, the combined method provided the accuracy increase of %7.26 for BDCom and %6.71 for QPct (Quality Percentage) when compared with the results of perceptual grouping method. The results achieved in this study demonstrate that the developed method is quite succesful in the extraction of buildings from color orthophoto and LiDAR point cloud data.