Real-World Validation of Drone Anomaly Detection Using Evidential Networks

In urban drone surveillance, conventional anomaly detection methods often falter amid unstable data and high noise levels. We propose a Sim2Real approach within a ValuationBased System framework that integrates expert knowledge into its model structure. Our method involves two phases: training on simulated drone trajectories to learn behavioral dependencies with probabilities, using a Dynamic Bayesian Network, and performing inference via an evidential network for explicit uncertainty modeling. The incorporation of expert-driven structural insights enhances the transfer of simulation-trained parameters to real-world conditions. Real-world tests with actual flight data demonstrate that our approach outperforms traditional techniques in detection accuracy and recall, robustly handling sensor noise and incomplete information.