This technical paper focuses on creating a failure-safe landing system for BLOVS UAVs, to achieve fully autonomous
drones which comply with the European Union Aviation Safety Agency regulations. This research tackled a specific part of a larger project, which aims to find a suitable segmentation model which can detect and avoid ground obstacles in the scenario of an emergency landing performed by an unmanned aerial vehicle. A deep learning approach is used, involving two segmentation architectures, U-Net and U-Net++, supported by different experiments to improve the performance of these models and in the end determine the best performing architecture. From the results, was concluded that using a segmentation approach is a suitable method to apply in this project, although a few limitations must be first settled to test this method in a real-life scenario.