Abstract

As a project for EE368, we investigated methods for reconstructing elevation data from a series of aerial images. First we explored a variety of methodologies for aligning the images and identifying the principal camera points for each photo. Once these absolute coordinates were obtained, we used epipolar geometry to extract topological data. These values were then inserted into a grid in an attempt to form a elevation map.

Introduction

Our goal is to explore methodologies for extracting elevation data from aerial images. We were given 74 images of rolling farmland, courtesy of Geovantage. The images were already corrected for camera distortion. We were not given any information concerning the camera or the images themselves.

In order to keep the problem tractable, we assumed several things about our data set. First, we made the reasonable assumption that all the pictures were taken with the camera facing straight down. To make our calculations possible, we also constrained the motion of the camera from one frame to the next to a translation in X and Y, assuming rotation to be negligible.

With these two assumptions, we developed the following methodology. First we manually registered the images by hand, finding the coordinate translation between one photo and the next. We then applied epipolar geometry to help us efficiently find the relation between pixels in the two frames. This information helped to limit the search area for our block matching algorithm, which found the best match for each pixel. The resulting difference in pixel location, combined with the our calculated change in camera location, was then used to evaluate a height for that particular pixel location.