Personal Protective Equipment Detection for Industrial Environments: A Lightweight Model Based on RTDETR for Small Targets

Abstract

Existing Personal Protective Equipment (PPE) detection research typically focuses on close-range scenarios, often neglecting small target detection in industrial surveillance. To address this gap, we propose LMD-RTDETR, a lightweight algorithm for small PPE targets. In the encoding stage of the neural network, we incorporated an Adaptive Inductive Frequency Learnable Position Encoding (AIFI-LPE) structure to enhance the model's ability to understand complex scenes. Additionally, the Dynamic Group Shuffle Transformer SlimNeck (DGST-SlimNeck) module and Multi-Path Spatial Semantic Feature Fusion (MP-SSFF) structure are optimized in the Neck network, enhancing the model's feature learning ability and achieving multi-scale feature fusion. These innovations significantly improve detection accuracy for small objects in complex scenes. We conducted extensive experiments on both a custom PPE dataset and the public VisDrone dataset. Compared to RT-DETR-r18 on the PPE dataset, LMD-RTDETR shows a 2.4% improvement in mean Average Precision at Intersection over Union thresholds from 50% to 95% (mAP@50:95). Simultaneously, it reduces parameters by 24.2% and computational complexity (Giga Floating Point Operations per Second, GFLOPs) by 6.8%. On the VisDrone dataset, it achieves an mAP@50 of 39.0%, demonstrating strong generalization capabilities. These results highlight LMD-RTDETR's effectiveness in small-target PPE detection within industrial settings, offering a balance between high accuracy and computational efficiency. Our work contributes to enhancing workplace safety through improved automated PPE detection, particularly in complex industrial environments with diverse monitoring conditions.