Bayesian Data Augmentation for DNN Retraining with Binomial Outcomes in Vision-Based UAV Landing

Abstract

In GPS-denied or cluttered urban environments, vision-based landing is essential for reliable UAV missions. Real-world landing sites are often unstructured and highly variable, requiring strong generalization by the perception system. Deep Neural Networks (DNNs) trained with synthetic data augmentation offer a scalable solution for learning landing-site features across diverse vehicle and environmental states. However, computationally expensive DNN retraining, along with challenging performance validation via test flights, limits exhaustive model fine-tuning and necessitates an optimized retraining pipeline. In this work, we deploy a Bayesian data augmentation framework integrated with a photorealistic simulator featuring high-fidelity vehicle dynamics to iteratively retrain the helipad detector DNN, maximizing landing performance as the objective function. We validate our framework with experiments in a photorealistic simulator under different environmental conditions and vehicle states, demonstrating improved landing performance and tighter confidence intervals on predicted landing outcomes.