Real-time YOLO across Cloud, edge, and IoT: architectures, optimisations, and deployment patterns

Abstract

Real-time object detection with the YOLO family is now deployed in cloud data centres, edge servers, and tiny IoT devices, each operating under different constraints of latency, bandwidth, memory, energy, and cost. In this paper, a deployment-centric survey of YOLO is presented, treating YOLO as a scalable family of models embedded in distributed systems rather than a single benchmarked network. First, object detection paradigms and the evolution of YOLO are reviewed, and a macro–meso–micro view of cloud, edge, and IoT deployment is introduced. Then, lightweight architectures and compression techniques—such as tiny and nano variants, efficient backbones, pruning, quantisation, and distillation—are surveyed, and their effects on accuracy, latency, model size, and energy are analysed. On the system side, deployment patterns are summarised, including cloud-centric serving, edge and fog deployments, IoT and tiny-device pipelines, and collaborative hierarchical inference with federated learning. Finally, open challenges and future directions for hardware-aware, deployment-aware YOLO co-design are outlined, and the potential of sustainable YOLO deployments to support UN Sustainable Development Goals in smart cities, healthcare, and environmental monitoring is highlighted.