Academic journal article URISA Journal

Automatic Generation of High-Quality Three-Dimensional Urban Buildings from Aerial Images

Academic journal article URISA Journal

Automatic Generation of High-Quality Three-Dimensional Urban Buildings from Aerial Images

Article excerpt

INTRODUCTION

Three-dimension building information is required for a variety of applications, such as urban planning, mobile communication, visual simulation, visualization, and cartography. Automatic generation of this information is one of the most challenging problems in photogrammetry, image understanding, computer vision, and GIS communities. Current automated algorithms have shown some progress in this area. However, some deficiencies still remain in these algorithms. This is particularly apparent in comparison to manual extraction techniques, which, although slow, are essentially perfect in accuracy and completeness.

Recent research covers extracting building information from high-resolution satellite imageries, high-quality digital elevation models (DEMs), and aerial images. For example, QuickBird and IKONOS high-resolution satellite imageries are used to acquire planemetric building information with one-meter horizontal accuracy (Theng 2006, Lee et al. 2003). However, aerial images are the primary source used to acquire accurate and reliable geospatial information. Lin and Nevatia (1998) proposed an algorithm to extract building wireframes from a single image. However, a single image does not provide any depth information. A pair of stereo images could also be used to extract building information (Avrahami et al. 2004, Chein and Hsu 2000). Using one pair of images is insufficient to extract the entire building because of hidden features that are not projected into the image pair.

Kim et al. (2001) presented a model-based approach to generate buildings from multiple images. Three-dimensional rooftop hypotheses are generated using three-dimensional roof boundaries and corners extracted from multiple images. The generated hypotheses then are employed to extract buildings using an expandable Bayesian network. Wang and Tseng (2004) proposed a semiautomatic approach to extract buildings from multiple views. They proposed an abstract floating model to represent real objects. Each model has several pose and shape parameters. The parameters are estimated by fitting the model to the images using least-squares adjustment. The algorithm is limited to parametric models only. In Shmid and Zisserman (2000), lines are extracted in the images and matched over multiple views in a pair-wise mode. Each line then is assigned two planes, one plane on each side. The planes are rotated, and the best-fitting plane is found. Planes then are intersected to find the intersection lines.

Henricsson et al. (1996) presented another system to extract suburban roofs from aerial images by combining two-dimensional edge information together with photometric and chromatic attributes. Edges are extracted in the images and aggregated to form coherent contour line segments. Photometric and chromatic contour attributes for adjacent regions around each contour are assigned to it. For each contour, attributes are computed-based on the luminance, color, proximity, and orientation, and saved for the next step. Contour segments then are matched using their attributes. Segments in three dimension are grouped and merged according to an initial set of plane hypotheses.

Fischer et al. (1997) extracted three-dimensional buildings from aerial images using a generic modeling approach that depends on combining building parts. The process starts by extracting low-level image features: points, lines, regions, and their mutual relations. These features are used to generate three-dimensional building part hypotheses in a bottom-up process. A step-wise model-driven aggregation process combines the three-dimensional building feature aggregates to three-dimensional parameterized building parts and then to a more complex building descriptor. The resulting complex three-dimensional building hypothesis then is back-projected to the images to allow component-based hypothesis verification. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.