Extraction and photogrammetric exploitation of features in digital images

Ashraf Nasr Sayed, Purdue University

Abstract

Three tasks are explored in this research: precise target location in a single digital image, matching between overlapping digital images, and the use of linear features in the photogrammetric reduction process. The use of least squares adjustment to precisely locate different features in digital images is explored. In particular, it is applied to determine the location of the center of a circular target and its radius. The mathematical model that describes a specific target in a digital image depends on: the measured quantities (gray levels), the ideal image function, and the point spread function (PSF). The gray level of a specific pixel is modeled as the convolution between the ideal image function and a point spread function. Two point spread functions, Gaussian and circ, are investigated. Several experiments with simulated data were conducted. The results indicate no significant difference between the accuracy obtained from the two functions. On the other hand, using the circ function, substantially reduces the time used in computations. In all experiments, errors of less than 0.1 pixel in the position of the center of the circle was obtained. Experiments were also performed to extract circular fiducial marks from aerial images. Accuracy of the inner orientation performed using the automatically extracted fiducial marks were comparable to those obtained when manual pointing was performed on an Analytical Plotter. A matching algorithm was developed based on a hierarchical approach with the following steps: (1) using an Interest Operator, select on the left photograph points which are likely match candidates; (2) perform unconstrained hierarchical matching using area correlation function; (3) perform relative orientation of the two overlapping photographs using the matched points; (4) perform constrained matching, along epipolar lines, to find corresponding points to the remaining unmatched points. Automatic relative orientation for two pairs of real images, one industrial and the other aerial, was performed to demonstrate the developed algorithm. Using linear features as primitives in the photogrammetric reduction process is an important step in digital photogrammetry. Several problems regarding straight lines and circular features were investigated. The main advantage of using features in the photogrammetric reduction is that no a priori information regarding corresponding image points are needed. Instead, only segments of corresponding features are required. Extensive experiments using several sets of simulated and real image data were performed. The results showed that linear features could be used as control for single photo resection and block adjustment. Relative orientation of a stereopair of images using straight lines is not geometrically possible, thus a relative orientation of a triplet of overlapping images using straight lines was performed. In both relative orientation and block adjustment, "pass features" are used in a role similar to "pass points".

Degree

Ph.D.

Advisors

Mikhail, Purdue University.

Subject Area

Civil engineering|Computer science

COinS