A Lightweight and Efficient Model for Safety Protection Equipment Detection Based on Improved RT-DETR

Safety helmets and reflective vests are critical protective gear for preventing accidents in agricultural greenhouse construction. To address the computational inefficiency of existing detection models on resource-constrained edge devices, this paper proposes an optimized RT-DETR framework featuring: (1) a GSConvNext-Net backbone that reduces parameter redundancy by 49.36%, (2) a CAS attention mechanism replacing self-attention in the AIFI module to enhance feature interaction, and (3) a lightweight-CCFM module strengthening cross-scale feature fusion in complex environments. Experimental validation on greenhouse safety datasets demonstrates 0.904 recall and 0.963 mAP@0.5 with 59.89% lower computational load than standard RT-DETR. When deployed on NVIDIA Jetson Xavier NX edge devices, the model achieves 32.2 FPS inference speed, enabling real-time safety compliance monitoring in confined greenhouse spaces while meeting practical deployment constraints.