A Multi-Stage Cascaded Inference Architecture for Real-Time Construction Safety Monitoring on Edge GPUs

Abstract

Computer-vision systems for construction safety sit at a design tension: the target device is an embedded GPU with tight latency and power budgets, yet the object space that makes a site truly readable spans personal protective equipment, heavy machinery, and handheld tools. A monolithic detector either shrinks to fit the hardware and loses rare-class accuracy, or grows to cover all categories and breaks real-time budgets. This paper proposes an architectural alternative: a family-specialized multi-stage cascade in which a coarse-grained primary detector triages the frame, and specialized secondary detectors handle disjoint object families within pre-qualified regions. The proposed design is implemented on a Volta-class embedded GPU through the DeepStream SDK, trained on a consolidated construction-safety dataset, and compared against representative edge and server baselines. The evaluation shows that the proposed cascade achieves real-time throughput at the edge while extending the detectable object space well beyond what prior edge work has achieved. It identifies the primary detector, rather than the specialized stages, as the dominant latency contributor and the most promising target for future optimization.