Image segmentation based emergency landing for autonomous and automated unmanned aerial vehicles

The growing application of unmanned aerial vehicles (UAVs) into diverse commercial and private applications entails serious risks and respectively requires regulations. The development towards autonomously flying drones enforces this need and justifies the development of additional safety levels. Therefore, this work presents an emergency landing operation which is able to identify safe landing spots for both the environment, as well as the UAV, using neural networks (NNs) to segment the image of a downward-facing camera. This image is projected onto the ground plane conserving metric scaling. The landing control is administered by a Behavior Tree (BT). Thereby, a fully autonomous emergency landing without any communication to the ground control, predefined safe landing spots or detailed maps with corresponding location is enabled. A software-in-the-loop (SIL) setup is implemented in which it is possible to conduct a safe emergency landing operation. Additionally, multiple real world test flights have been conducted gathering data to evaluate this method.