Internet of Things, in particular, the concept of Industrial Internet of Things (IIoT), is one of the key technological pillars of the Fourth Industrial Revolution, also known as Industry 4.0. In this context, one of the areas of interest is safety, whereby multiple intelligent sensors may be permanently connected to a central system to autonomously or semi-autonomously identify safety hazards. Vision systems are a popular sensor in the safety domain as they can simultaneously monitor many different safety concerns. However, the continuous video stream transmission and the increasing number of intelligent devices in IIoT networks introduce additional pressure on the network. There is a risk that the network resources may become overloaded. This paper proposes and discusses a reference architecture for identifying safety risks. The architecture allows multiple sensors to be plugged into the system. The input of the different sensors is then dynamically weighed as the risk situation evolves. The architecture explores sensor-level intelligence (at the edge layer) to mitigate the network overloading problem. Edge agents quickly assess the risk, deciding whether or not to forward their signals to a local cloud agent for further processing. The cloud agent can then selectively request more information from other edge agents. The architecture is tested in a use case for operators’ safety in the assembly of aircraft components and uses intelligent vision systems as safety devices. In the selected use case, the accuracy of the system and its impact on the network load are assessed.