CR-DETR: an efficient model for the personal protective equipment detection in construction scenarios

Abstract In the process of building construction, personal protective equipment (PPE) serves as the final safeguard for worker safety, and by monitoring its wearing status in real time, the likelihood of casualties can be significantly lowered. However, the construction scene is usually complex and dynamic, the targets of workers’ protective equipment are small and difficult to detect, and the computational resources of the edge equipment of the deployment algorithm are limited. To solve these problems, this study proposes CR-DETR, an improved detection algorithm based on RT-DETR networks. First, parameter reduction is achieved by using the lightweight MobileNetV4 as the backbone network. Second, a feature interaction module Clo-AIFI based on dual-branch structure is designed, and the BiFPN structure with P2 layer is introduced to enhance the ability of fusing multi-scale features. Finally, the dynamic upsampler DySample is used to replace the nearest neighbor upsampling of the original model to improve the feature expression efficiency. According to the experimental outcomes, the accuracy of the model is 98.5%, the number of parameters is only 11.73 million, the GFLOPs is 46.1, and the detection speed is 86.1 FPS. Compared with the benchmark model, this model achieves significant lightweighting while maintaining high detection accuracy, and can meet the edge deployment requirements of protective equipment detection on construction sites.