DTM generation from digitized aerial photos of a complex scene by employing pattern recognition with the Fourier transform and multiresolutional feature-based stereo matching

The main object of this research has been to generate digital terrain elevation data from a stereopair of digitized aerial photographs. To achieve this, a systematic approach was developed. This approach involves (1) preliminary preparation of the digitized photographs using pattern recognition techniques, in particular, with the Fourier transform; (2) performing multiresolutional, feature-based stereo matching on the stereo pair. Photogrammetric techniques were then used to calculate the terrain elevation data using coordinates derived from the stereo pair.;Pattern recognition by employing the Fourier transform was successful in locating the positions of fiducial marks in the aerial photos. Multiresolutional, feature-based stereo matching was implemented by applying geometrical and similarity constraints. Laplacian and Gaussian filters with different sizes and Mallat's wavelet transform were used to decompose the original images into a series of images with different scales. The terrain elevation data collected from the developed system with different multiresolution approaches was subsequently compared with data collected by manually operating the analytical stereo plotter, Kern DSR-14. The results show that the elevation data collected by the developed approach is very similar to the data collected by manually operating the stereo plotter.;Compared with the digital terrain model generated by manually operating the stereo plotter and the digital terrain models generated by the developed approach, it was found out that the developed approach can successfully generate a digital terrain model from a given stereo pair. The C factor of the developed system is around 2000 for the test images, and is very close to the C factors of analytical stereoplotters. During experiments, it was discovered that when an area is covered with many trees and bushes, it is very difficult to locate the stereo correspondences from the given stereo pair. The developed system still has difficulty to solve this problem. Two image pyramids, Laplacian and Gaussian filter with different sizes and Mallat's wavelet, were used to create a series of images with different scales. Then, the stereo matching system developed in this research was applied to generate digital terrain model from different image pyramids. It was found out that the digital terrain models generated by applying different image pyramids were very similar. However, Mallat's image pyramid has faster processing speed than that of Laplacian and Gaussian image pyramid.