Using Algorithm Selection for Adaptive Vehicle Perception Aboard UAV

Abstract

Surveillance sensors aboard UAV are affected by environmental influences, e.g. atmospheric or topographic factors. This paper proposes a method for the automatic adaption of airborne sensor applications such as street surveillance to changing environmental conditions, preventing overall performance degradation with minimum human intervention. The basic principle of the concept relies on the selection of the most appropriate data processing algorithm available on board. To facilitate the determination of the most effective algorithm, performance models are used to predict the expected suitability of each algorithm for the given environmental conditions. Modeling the relation between the environmental state and the performance of the algorithms is achieved by two approaches leveraging expert knowledge and machine learning methods. An evaluation was carried out in simulation as well as in real flight experiments showing that the proposed method is able to improve overall vehicle perception performance.