Image Orientation by Fitting Line Segments to Edge Pixels
Ebner and Strunz (1988) proposed another approach to recovering the exterior orientation by Digital Elevation Model (DEM) grid patches. Measuring more than three points on the same patch will provide a set of plane equations. The exterior orientation can be determined when all of the measured planes fit to the corresponding control surfaces by minimizing their Z-axis differences. This method is not suitable for the surfaces are near vertical. Jaw (1999) improved the method by minimizing the difference along the normal direction of the surface instead of the Z- axis. These area-based methods are suitable while the DEM has been known or can be derived from other sources like Light Detecting And Ranging (LiDAR) or Synthetic Aperture Radar (SAR) data, but it requires more computations and still needs manual measurements.
In this research, the building model is adopted as the control element for solving image exterior orientation. The wire -frame model of the known building is decomposed into 3D line segments and projected onto the image based on the approximate orientation. The optimal image orientation is solved by varying orientation parameters to achieve the optimal fit between projected line segments and extracted edge pixels on the image. This solution applies least-squares fitting algorithm and does not require any image measurement.
The general idea of the least-squares fitting is to minimize the squares summation of the distances from each edge pixel to the corresponding projected line segment. The fitting algorithm starts from a given first approximation of image orientation and solves for the correction of image orientation iteratively to find the optimal fitting. A buffer of each projected line segment is provided to determine that each pixel is corresponding to which line segm ent. It is assumed that the interior orientation of the camera is known, so that the unknown parameters are the coordinates of the perspective center and 3 rotation angles. The study cases include a close-range image pair and an aerial photo. The accuracy of image orientation by fitting line segments to edge pixels will be analyzed.
2. Orientation Recovery using Linear Features
Recovering the image orientation refers to restoring the geometric relation between the sensor and the object at the moment of taking photograph, which is the fundamental task of the photogrammetric, remote sensing or computer vision applications. It is usually divided into two parts as consideration, the interior orientation and the exterior orientation. The exterior orientation refe rs to the geometric relation of the sensor in the object space, which can be represented by parameters such as the object coordinates (X0, Y0, Z0) of the perspective center, and the rotation angles of the sensor. The exterior orientation parameters can be solved by space resection for single photo, independent model adjustment or bundle adjustment for multiple photos. These traditional techniques are based on measuring points on the images and corresponding them to their 3D coordinates in object space. However, the point- based methods are difficult to be automated due to the lack of robust image interpretation and intelligent
point- matching algorithm. The tremendous points-measuring work is still remains for human.
