Generative AI-enhanced UAV flight simulation to accelerate UAV deep learning research

Unmanned Aerial Vehicles (UAVs) are powerful flexible monitoring devices, and their data is increasingly used to train deep learning models for various important applications such as monitoring sensitive structures or search-and-rescue operations. However, local regulations or weather conditions often restrict how often and where a UAV can be flown. For example, for road inspections, there is a specific safety buffer that the UAV cannot enter, limiting the camera angle's perspective. In addition, for specific applications, such as damage detection, the object of interest (for example, potholes) cannot always be found within the environment, leading to data scarcity and unbalanced datasets.

To mitigate the problem of unbalanced datasets, to increase the dataset size and improve the performance of deep learning models, an increase in the usage of synthetic datasets can be seen. The advantage of simulated datasets is that a range of flight paths, flight telemetries, sensors (optical and hyperspectral) and environments can be mimicked in a highly realistic manner. Examples of such datasets can be found for autonomous driving (Virtual KITTI), 3D object recognition (PASCAL). Simulated datasets for UAV applications are rare. The most notable one is Midair (https://midair.ulg.ac.be/) [1].