Deploying a Feedback Loop-based Training Strategy for Deep Learning-based Drone Detection

Tamara R. Lenhard, Andreas Weinmann, Stefan Jäger, Eva Brucherseifer

Abstract: Detecting drones in real-world scenarios with high reliability (e.g., for protecting critical infrastructures) is an essential yet challenging computer vision task due to the intricate and continuously evolving nature of drone technology. In this paper, we consider a feedback loop-based training strategy to address the need for robust drone detection systems. Leveraging game engine-based simulations within three-dimensional environments, our approach facilitates the application-oriented refinement of synthetic training data in an iterative manner, effectively narrowing the simulation-reality gap. By incorporating a small amount of real-world data into the training process, our strategy demonstrates its efficacy across multiple real-world datasets, surpassing the performance of models derived via zero-shot sim-to-real transfer learning. Our findings highlight the practical relevance of this approach, especially in surveillance settings, and emphasize its potential to enhance deep learning models for drone detection.