Real-time safe validation of autonomous landing in populated areas: from virtual environments to Robot-In-The-Loop
Hector Tovanche-Picon,Javier González-Trejo,Ángel Flores-Abad,Miguel Ángel García-Terán,Diego Mercado-Ravell
DOI: https://doi.org/10.1007/s10055-024-00965-6
IF: 4.697
2024-03-06
Virtual Reality
Abstract:Safe autonomous landing for Unmanned Aerial Vehicles (UAVs) in populated areas is a crucial aspect for successful integration of UAVs in populated environments. Nonetheless, validating autonomous landing in real scenarios is a challenging task with a high risk of injuring people. In this work, we propose a framework for safe real-time and thorough evaluation of vision-based autonomous landing in populated scenarios, using photo-realistic virtual environments and physics-based simulation. The proposed evaluation pipeline includes the use of Unreal graphics engine coupled with AirSim for realistic drone simulation to evaluate landing strategies. Then, Software-/Hardware-In-The-Loop can be used to test beforehand the performance of the algorithms. The final validation stage consists in a Robot-In-The-Loop evaluation strategy where a real drone must perform autonomous landing maneuvers in real-time, with an avatar drone in a virtual environment mimicking its behavior, while the detection algorithms run in the virtual environment (virtual reality to the robot). This method determines the safe landing areas based on computer vision and convolutional neural networks to avoid colliding with people in static and dynamic scenarios. To test the robustness of the algorithms in adversary conditions, different urban-like environments were implemented, including moving agents and different weather conditions. We also propose different metrics to quantify the performance of the landing strategies, establishing a baseline for comparison with future works on this challenging task, and analyze them through several randomized iterations. The proposed approach allowed us to safely validate the autonomous landing strategies, providing an evaluation pipeline, and a benchmark for comparison. An extensive evaluation showed a 99% success rate in static scenarios and 87% in dynamic cases, demonstrating that the use of autonomous landing algorithms considerably prevents accidents involving humans, facilitating the integration of drones in human-populated spaces, which may help to unleash the full potential of drones in urban environments. Besides, this type of development helps to increase the safety of drone operations, which would advance drone flight regulations and allow their use in closer proximity to humans.
computer science, interdisciplinary applications, software engineering,imaging science & photographic technology