Mustererkennung zur Detektion von Rehkitzen in Thermalbildern

Abstract

During pasture mowing 25% of the newborn fawns are killed [Jar02]. They are born in the meadows in the months May and June, the time of pasture mowing. Because of their innate instinct to remain motionless on the ground in case of imminent danger, they are often overlooked by farmers and inadvertently killed during mowing. This creates the demand for systems capable to detect these hidden animals. The paper presents a functional model of such a system, consisting of an airborne platform (octocopter) equipped with a thermal camera and a miniature computer. While the octocopter flies GPS-controlled search missions above the field, thermal images are taken permanently. A pattern recognition algorithm scans each thermal image captured during the flight in order to automatically detect fawns. The algorithm is based on a compression method. During the compression of an image a dictionary is created containing only non-redundant information. The dictionaries of the images recorded are compared with those of the training set categorised by the classes of fawns and meadows. The tested image is assigned to the class whose dictionaries exhibit greatest similarity. The decision, whether a fawn is recognized or not, is based upon this assignment.