Today's advances in embedded technologies enable sensor image processing to be performed onboard mini or micro UAVs. Recent developments have shown that classification in real time based on self-trained computer vision algorithms shows promising results. However self-trained algorithms are based on considerable amounts of training data that is not easily accessible for airborne applications due to high costs and complicated setups of real life flight This paper proposes the usage of a synthetic environment to generate high amounts of training data allowing prototyping of computer vision algorithms. It discusses the use of ITEM, a test bed designed to use a virtual environment for simulation of sensor output. Furthermore a self-trained algorithm, used for vehicle detection in real time, is introduced. The example algorithm is then validated and achieved results are discussed, leading to an evaluation concept for computer vision algorithms to determine real world relevance. The proposed method may lead to more cost efficient research and development of such algorithms as well as increase in detection quality of classifiers.
展开▼
