Abstract:
The demand for an increased degree of automation of unmanned aerial vehicles also affects the operation and functionality of the on-board sensor systems for environmental sensing. Such systems serve on the one hand as navigational support (e.g. sense & avoid), but on the other hand primarily as payload for conducting reconnaissance and surveillance missions. The performance of these systems is influenced by varying environmental conditions, e.g. atmospheric conditions or topographic changes. Against this background, this thesis proposes a method for adaptive airborne sensor deployment, enabling a response to changes in environmental conditions and thus counteracting degradation of system performance without a continuous intervention by an operator. The basic methodology of adaptive sensor deployment is based on an algorithm selection which uses models for the performance behavior of the various algorithms available on board for sensor data processing. The algorithms use different methods for vehicle recognition on sensor data of imagery sensors. In order to determine the performance, environmental states were used to describe the situational environment. The modelling of the relationship between the environmental state and the algorithm performance is based on expert knowledge on the one hand and on methods of machine learning on the other. For the latter, a simulation environment was used to collect corresponding data. The implemented system for adaptive sensor deployment was evaluated by experiments in the simulation environment as well as by flight experiments. It was found that an improvement of the reconnaissance performance could be achieved by algorithm selection compared to the sole use of the most powerful algorithm available for object recognition. Furthermore, it was shown that a synthetic sensor simulation can be used to generate the models for performance determination and at the same time be usable in real scenarios. It turned out that the accuracies of the performance determination are comparable in simulated and real situations. Finally, the implemented adaptive sensor deployment could be successfully demonstrated in real flight experiments.