Real-time computer vision for safety and efficiency in smart infrastructure

This systematic review summarises the current state of real-time computer vision practice for smart infrastructure, and compiles evidence on efficiency and safety outcomes. In accordance with the PRISMA 2020 guidelines, we preregistered a protocol and conducted a search of five bibliographic databases for studies published between January 2015 and 24 September 2025. Nineteen field-validated studies met the eligibility criteria. Across the wastewater, buildings, structural assets domains and ports, transport, two recurring value streams were observed: (i) safety gains that shorten the event-to-intervention interval (e.g. detection of trespassing or incidents, vision-based metrology for displacement or vehicle load, and proximity risk cues), and (ii) efficiency gains that improve throughput, energy performance or process stability (e.g. demand-responsive building control, station and intersection analytics, wastewater foam segmentation and digital twin–linked crane operations). Most systems use technical solutions such as semantic segmentation for state/surface estimation, detector–tracker pipelines, or vision-based metrology, edge-cloud, implemented via edge-first, edge–cloud, or on-premises/platform-integrated deployments. However, cross-study comparability remains limited because many papers do not consistently report sustained throughput (FPS), end-to-end latency, network/hardware context and energy consumption. To address this, we present a unified implementation and governance framework, along with a reporting checklist. This framework ties model accuracy to operational decision pathways, access/privacy controls and real-time constraints, and lifecycle monitoring.