Geolocation of photographs by means of horizon matching with digital elevation models

Geolocation is the process of assigning geographic coordinates to photographs. It is important in counter terrorism, photo tourism, community remote sensing, and robot navigation. Prior work has focused on matching photographs against a database of already geolocated photographs. This project adopts a different approach, extracting a horizon curve from the photograph and searching for the best match with horizon curves generated from digital elevation models. A wide variety of photographs were tested from eastern and western regions of the U.S. They included landscapes, beach scenes, and even indoor scenes with the horizon visible through a window. The horizons differed in length, orientation, distinctiveness, visibility, and vegetation cover. The geographic search regions were extensive and covered over 500,000 square kilometers. Of 100 photographs tested, the algorithm successfully geolocated 83%, with a mean error of only 300 meters. It was learned that successful geolocation must account for the curvature of the earth, atmospheric refraction, and terrain elevation inaccuracies. Implemented in Java on a PC, the execution time is proportional to the search area and averages 7600 square kilometers per minute.