Ground-Based UAV Detection on Edge Platforms: A Survey of Architectures, Datasets, and Hardware for Deployment

Abstract

Deploying a UAV detector on edge hardware requires choosing an architecture, a quantization path, and a platform together, yet no prior work addresses these decisions jointly. This survey addresses all three for ground-based counter-UAV detection, where a fixed or slow-moving camera detects flying UAVs in real time within a 60 W power envelope. We review 35 detection models across five architecture families benchmarked on COCO val2017, with deployment measurements collected for 14 models across five edge platforms ranging from 9 to 60 W. Five ground-based RGB datasets, selected primarily for capture geometry and characterized across six deployment-relevant axes, form the training and evaluation pool. Four findings stand out. First, YOLO-family CNNs dominate the accuracy–efficiency frontier: eight models form the strict Pareto frontier; of the 23 models within 3 pp of the frontier, 14 have direct INT8 edge-measurement evidence, all YOLO-CNN. Second, INT8 quantization loss is small at a median of 1.15 pp across 42 model–platform pairs, but toolchain-dependent. Third, prior object-detection literature reports up to 20 pp accuracy loss from domain shift without adaptation, larger than the quantization loss, both drawn from non-UAV datasets. Fourth, hardware acquisition dominates three-year total cost of ownership across all six platforms, with electricity contributing only 9–24% at U.S. average electricity prices. We contribute a challenge-to-metric evaluation framework; a five-stage deployment guide and cross-dataset training protocol, both recommendations, not validated results; and a discussion of four open problems for practitioners.