Autonomous unmanned aircraft need a good semantic understanding of their surroundings to plan safe routes or to find safe landing sites, for example, by means of a semantic segmentation of an image stream. Currently, Neural Networks often give state-of-the-art results on semantic segmentation tasks but need a huge amount of diverse training data to achieve these results. In aviation, this amount of data is hard to acquire but the usage of synthetic data from game engines could solve this problem. However, related work, e.g., in the automotive sector, shows a performance drop when applying these models to real images. In this work, the usage of synthetic training data for semantic segmentation of the environment from a UAV perspective is investigated. A real image dataset from a UAV perspective is stylistically replicated in a game engine and images are extracted to train a Neural Network. The evaluation is carried out on real images and shows that training on synthetic images alone is not sufficient but that when fine-tuning the model, they can reduce the amount of real data needed for training significantly. This research shows that synthetic images may be a promising direction to bring Neural Networks for environment perception into aerospace applications.

State-of-the-art object detection models need large and diverse datasets for training. As these are hard to acquire for many practical applications, training images from simulation environments gain more and more attention. A problem arises as deep learning models trained on simulation images usually have problems generalizing to real-world images shown by a sharp performance drop. Definite reasons and influences for this performance drop are not yet found. While previous work mostly investigated the influence of the data as well as the use of domain adaptation, this work provides a novel perspective by investigating the influence of the object detection model itself. Against this background, first, a corresponding measure called sim-to-real generalizability is defined, comprising the capability of an object detection model to generalize from simulation training images to real-world evaluation images. Second, 12 different deep learning-based object detection models are trained and their sim-to-real generalizability is evaluated. The models are trained with a variation of hyperparameters resulting in a total of 144 trained and evaluated versions. The results show a clear influence of the feature extractor and offer further insights and correlations. They open up future research on investigating influences on the sim-to-real generalizability of deep learning-based object detection models as well as on developing feature extractors that have better sim-to-real generalizability capabilities.

A key necessity for the safe and autonomous flight of Unmanned Aircraft Systems (UAS) is their reliable perception of the environment, for example, to assess the safety of a landing site. For visual perception, Machine Learning (ML) provides state-of-the-art results in terms of performance, but the path to aviation certification has yet to be determined as current regulation and standard documents are not applicable to ML-based components due to their data-defined properties. However, the European Union Aviation Safety Agency (EASA) published the first usable guidance documents that take ML-specific challenges, such as data management and learning assurance, into account. In this paper, an important concept in this context is addressed, namely the Operational Design Domain (ODD) that defines the limitations under which a given ML-based system is designed to operate and function correctly. We investigated whether synthetic data can be used to complement a real-world training dataset which does not cover the whole ODD of an ML-based system component for visual object detection. The use-case in focus is the detection of humans on the ground to assess the safety of landing sites. Synthetic data are generated using the methods proposed in the EASA documents, namely augmentations, stitching and simulation environments. These data are used to augment a real-world dataset to increase ODD coverage during the training of Faster R-CNN object detection models. Our results give insights into the generation techniques and usefulness of synthetic data in the context of increasing ODD coverage. They indicate that the different types of synthetic images vary in their suitability but that augmentations seem to be particularly promising when there is not enough real-world data to cover the whole ODD. By doing so, our results contribute towards the adoption of ML technology in aviation and the reduction of data requirements for ML perception systems.