Trade-Off Analysis of YOLO Architectures for Embedded UAV-Based Transmission Line Inspection

Abstract

Abstract This paper presents a comprehensive evaluation of YOLO-based architectures for automated transmission tower inspection using UAV imagery under edge computing constraints. The study investigates the performance of different model variants in a resource-constrained embedded environment, focusing on the trade-offs between detection accuracy, inference latency, and computational efficiency. A key contribution of this work is the systematic analysis of data augmentation under limited-data conditions, using a specialized proof-of-concept dataset of 352 annotated UAV images, which constrains broad statistical generalization. The results demonstrate that data augmentation consistently improves recall, enhancing detection sensitivity, but also reduces localization accuracy as measured by $$\text {mAP}_{50-95}$$ , revealing a critical trade-off between detection completeness and spatial precision. In addition, the proposed approach is validated through real-time deployment on a Luxonis OAK-D S2 embedded vision platform. The experiments show that intermediate models such as YOLOv8s achieve a strong balance between performance ( $$\text {mAP}_{50} = 0.9038$$ ) and latency. In contrast, lightweight models achieve inference times as low as 147 ms while maintaining competitive $$ F_1$$ Scores, making them suitable for real-time UAV inspection. The results further indicate that increasing model complexity does not necessarily yield proportional gains in detection performance under embedded constraints, underscoring the importance of hardware-aware optimization. Overall, this work provides practical guidelines for selecting and deploying object detection models in real-world inspection systems, contributing to the development of efficient and reliable solutions for power infrastructure monitoring.