Edge-Ready Multimodal and Geometry-Aware Perception for UAV-Based High-Voltage Power Line Inspection: A Systematic Review

Abstract

Unmanned Aerial Vehicles (UAVs) have become a promising alternative for high-voltage power line inspection because they reduce operational risk, inspection time, and human exposure to hazardous environments. However, reliable real-time power line perception remains a critical bottleneck for autonomous deployment, particularly when thin cable structures must be detected under limited onboard computation, adverse illumination, occlusion, motion blur, and sensor constraints. This systematic review analyzes UAV-based power line detection and segmentation methods from the perspective of deployable visual perception. Following the PRISMA methodology, 104 studies published between 2021 and 2026 were selected from Scopus. The included studies report quantitative evaluation metrics and address classical vision or deep learning-based power line perception, while works focused exclusively on tower inspection, route planning, or non-visual inspection were excluded. A four-category taxonomy is proposed according to operational constraints: efficiency-oriented methods (32.7%), accuracy-oriented detection (15.4%), sensor-fusion approaches (23.1%), and shape/geometry-aware methods (28.8%). The results show that YOLO-based models dominate edge-oriented scenarios, with reported performance reaching up to 98.91 FPS and 97.2% mAP, while emerging RT-DETR and Mamba-based architectures are narrowing the performance gap. Infrared sensing improves robustness under adverse conditions, outperforming RGB by up to 6.4 mAP points in controlled comparisons, whereas multimodal knowledge distillation offers a promising path toward lightweight RGB-only deployment. Geometry-aware representations, including keypoints, segmentation masks, Hough-domain features, and topology-aware models, are increasingly relevant because conventional bounding boxes are poorly suited to elongated cable structures. The review also identifies major methodological limitations: 85.6% of the studies rely on private datasets, reported metrics are highly inconsistent, and only a small fraction validate their models on real embedded hardware. Future research should prioritize standardized public benchmarks, energy-aware edge evaluation, multimodal-to-unimodal distillation, and closed-loop perception-navigation systems. This review provides a practical roadmap for selecting UAV-based power line perception methods according to hardware capacity, sensing modality, environmental conditions, and real-time deployment requirements.