Filling the Gaps: Using Synthetic Low-Altitude Aerial Images to Increase Operational Design Domain Coverage

Area of Science:

• Aerospace Engineering
• Computer Science
• Artificial Intelligence

Background:

• Safe and autonomous flight of Unmanned Aircraft Systems (UAS) relies on robust environmental perception.
• Machine Learning (ML) offers advanced visual perception but faces certification hurdles due to data-dependent properties.
• European Union Aviation Safety Agency (EASA) provides guidance on ML challenges like data management and learning assurance.

Purpose of the Study:

• To investigate the utility of synthetic data in complementing real-world datasets for ML-based visual object detection.
• To assess if synthetic data can increase the Operational Design Domain (ODD) coverage for ML systems.
• To evaluate synthetic data generation techniques for improving ML model performance in aviation.

Main Methods:

• Generated synthetic data using EASA-recommended methods: augmentations, stitching, and simulation environments.
• Augmented a real-world dataset with synthetic data to train Faster R-CNN object detection models.
• Focused on detecting humans on the ground for landing site safety assessment.

Main Results:

• Synthetic data generation techniques vary in suitability for increasing ODD coverage.
• Augmentations proved particularly promising for enhancing datasets with limited real-world data.
• The study provides insights into generating and utilizing synthetic data for ML in aviation.

Conclusions:

• Synthetic data can effectively augment real-world datasets to improve ODD coverage for ML perception systems.
• This research contributes to the adoption of ML in aviation by addressing data requirements.
• The findings support the development of certifiable ML-based systems for UAS applications.