Proceedings Article | 13 June 2024
KEYWORDS: Target detection, Object detection, Head, Safety, Detection and tracking algorithms, Small targets, Feature extraction, Data modeling, Performance modeling, Neck
To address the challenges of low detection accuracy and substantial model parameter size present in traditional safety helmet detection algorithms, this study introduces an enhanced algorithm known as P2DFE-YOLOv8n, specifically crafted for the detection of safety helmet usage. This algorithm builds upon the foundation of YOLOv8n by incorporating a specialized detection head for small targets, aimed at improving the accuracy of safety helmet detection. Simultaneously, it eliminates the detection head for large targets to reduce the model's parameter count and computational complexity, without compromising detection accuracy. The inclusion of the FasterNet Block and C2f modules in the model architecture serves to further streamline the model and enhance its accuracy. Additionally, the employment of the EMA (Efficient Multi-Scale Attention) attention mechanism within the neck networks allows for focused processing of crucial information related to the individual wearing the helmet, thereby significantly increasing detection precision. Experimental evaluations conducted on the proprietary Hhelmet safety gear database demonstrate that the refined P2DFE-YOLOv8n algorithm achieves an mAP@0.5 score of 92.3%, with a remarkably reduced parameter volume of just 0.76M. Compared to the original YOLOv8n model, the refined algorithm presents a reduction in parameter size and model size by 2.25M (74.8%) and 4.294MB (70.6%), respectively. Furthermore, the enhanced model showcases improvements in accuracy, recall rate, mAP@0.5, mAP@0.5:0.95, and FPS by 1.8%, 0.2%, 1.0%, 0.2%, and 14, respectively, underscoring the efficacy of the proposed modifications in optimizing both performance and efficiency in safety helmet detection tasks.
