FY26 Faculty CAT-A Proposal: Accelerating AI Adoption in Aviation Safety: A Benchmark Dataset for Uncrewed Aircraft Systems (UAS) Detection Using Diffusion Models

The integration of Artificial Intelligence (AI) into aviation safety-critical systems is urgently needed to address escalating risks of mid-air collisions (MACs) between manned aircraft and Uncrewed Aerial Systems (UAS). Recent accidents—including a drone-firefighting helicopter collision in California and the Airline Airlines 5342 crash at DCA airport—underscore the inadequacy of current detection systems and the critical lack of diverse, high-fidelity benchmark datasets for training and validating machine learning (ML) algorithms. This project addresses this gap by developing a novel, open-access benchmark dataset for UAS detection, leveraging advanced generative models to overcome the cost, safety, and data scarcity challenges inherent to real-world aviation data collection.
While existing research has explored Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) for synthetic data generation, this study pioneers the use of diffusion models to create photorealistic, scenario-specific UAS images under varying environmental conditions (e.g., weather, lighting, occlusion). A comparative analysis of generative techniques will be conducted to optimize dataset diversity and fidelity, ensuring robust training of state-of-the-art (SOTA) ML models.
By bridging the gap between generative AI and aviation safety, this research will accelerate the deployment of reliable UAS detection systems, mitigating collision risks while establishing a replicable framework for AI benchmarking in safety-critical domains.