Design of an improved YOLOv9-driven real-time multiobject detection system for construction site safety monitoring
TL;DR
An improved YOLOv9 real-time target detection system architecture based on edge-cloud collaboration is proposed, providing a powerful system-level solution for the large-scale, low-cost deployment of smart construction site safety monitoring networks.
Abstract
Addressing the engineering challenges of multi-scale target occlusion, high miss rates for small targets, and severely limited computing power of edge devices in unstructured construction site environments, this paper proposes an improved YOLOv9 real-time target detection system architecture based on edge-cloud collaboration. By reconstructing the hardware-friendly PGI-Lite lightweight backbone network, operator redundancy and memory jitter in feedforward inference are significantly reduced. An innovative high-frequency spatial attention mechanism adaptively enhances the response to abrupt edge changes, supplemented by dynamic asymmetric cross-union ratio loss to eliminate regression oscillations, thoroughly optimizing the feature mapping logic under dense working conditions. Real-world test results on actual edge computing devices demonstrate that, with a sharp reduction in the number of parameters to 5.8M, the average accuracy is improved by 2.7% to 89.5%, and the inference throughput reaches 132 FPS. This perfectly achieves an optimal balance between computing power and accuracy in all-weather industrial monitoring scenarios, providing a powerful system-level solution for the large-scale, low-cost deployment of smart construction site safety monitoring networks.