Human detection in UAV imagery using deep learning: a review

Deep learning models have enabled real-time human detection in Unmanned Aerial Vehicle (UAV) imagery for various applications, such as autonomous navigation and surveillance. Despite recent advances, detecting small objects—such as humans in UAV imagery—remains challenging. Although research activity has increased over the past few years, no literature review specifically focused on human detection in UAV imagery has been identified. Given the complexity and importance of this task, this study analyzes articles published over the last five years to address this gap. A synthesis of the principal detection methods, the specific challenges of small object detection, and the corresponding evaluation metrics are presented. Unlike previous review articles, this paper identifies and discusses original research papers focused on human detection in UAV imagery, as well as the major UAV datasets containing annotated persons. Although small object detection remains challenging for single-stage models, “You Only Look Once” (YOLO) architectures are predominantly adopted in UAV human detection research (70% of the reviewed original articles employ a YOLO model or one of its variants). The primary bottlenecks identified for accurate human detection in UAV imagery involve the limited availability of large, properly annotated datasets, leading 48% of the studies to rely on bespoke datasets, and the integration of optical and thermal imagery, which is addressed in 15% of the research. Cloud-based solutions and the deployment of lightweight, fast models on mobile devices using embedded boards emerge as the key directions for future work.