Aerialguard: Edge-Deployed Zero-Shot Anomaly Detection for UAVs

Abstract

Autonomous UAV patrol requires real-time perception capable of identifying novel visual anomalies without prior labeled examples. Closed set detectors fail on unseen objects, while CLIP-based anomaly methods designed for ground-level imagery have not been validated for the aerial domain. We present AerialGuard, an end-to-end zero-shot visual anomaly detection system deployed entirely on a real autonomous UAV. AerialGuard combines a domain-aligned visual encoder finetuned on aerial imagery via contrastive learning, a Mahalanobis one-class scorer trained solely on normal examples, and a deterministic behavior router that maps anomaly confidence directly to closed-loop flight actions. A key finding is that contrastive finetuning creates a polarity inversion in embedding space, normals drift farther from the centroid than anomalies, causing Euclidean scoring to fail, which Mahalanobis whitening corrects. The scorer achieves 0.9968 AUROC on HERIDAL, a geographically disjoint alpine benchmark, with no anomaly training examples, a single threshold fitted on normal VisDrone crops transfers without modification across all evaluation conditions. The full pipeline runs at 6.4 FPS on a Jetson Orin Nano during autonomous flight. A five-flight validation at three altitudes confirms zero false state transitions and correct anomaly triggering, demonstrating that zero-shot visual scoring can drive reliable closed-loop UAV behavior under edge resource constraints.