基于YOLOv8n的施工场景下安全帽佩戴检测算法  

Detection Algorithm for Helmet Wearing in Construction Scenarios Based on YOLOv8n

作　　者：盛鹏 张敏 晋从乾 朱子玄 江文豪 SHENG Peng;ZHANG Min;JIN Cong-qian;ZHU Zi-xuan;JIANG Wen-hao(School of Artificial Intelligence and Big Data,Hefei University,Hefei 230601,China)

机构地区：[1]合肥大学人工智能与大数据学院,安徽合肥230601

出　　处：《计算机技术与发展》2025年第3期34-39,共6页Computer Technology and Development

基　　金：安徽省高校优秀科研创新团队项目(2022AH010095);合肥大学人才基金项目(20RC19)。

摘　　要：在施工场景下,针对安全帽检测任务中存在小目标较多、目标易被遮挡和易受环境因素干扰等问题,提出一种改进YOLOv8n的安全帽佩戴检测算法。首先,设计一种双向特征金字塔网络,通过跳跃连接增强特征表达能力,优化特征融合,并增加160×160有效特征层来提高小目标的检测精度;其次,利用分组卷积的思想和共享参数的策略重构检测头模块,在保证检测精度不受影响的前提下减少模型的参数量;最后,引入基于大型可分离核注意力(LSKA)和瓶颈注意力模块(BAM)设计的注意力引导模块,将其融入到空间金字塔池化(SPPF)模块中,该模块通过LSKA扩大感受野,提升模型对遮挡目标的局部特征捕捉,同时通过BAM关注空间和通道,过滤环境噪声,减少环境干扰对模型检测的影响。实验结果表明,改进后的算法相较于基线算法,准确率提高了0.2百分点,召回率提高了2.1百分点,mAP@0.5提升了2.6百分点,此外,参数量仅为2.0 M,FPS达到了84.5帧。To address the issues of a significant number of small-sized targets,susceptibility to occlusion,and environmental interference in the task of safety helmet detection in construction scenarios,an improved YOLOv8n-based safety helmet detection algorithm is proposed.Firstly,a bidirectional feature pyramid network(BiFPN)is designed to enhance feature expression through skip connections,optimize feature fusion,and add a 160×160 effective feature layer to further improve the detection accuracy of small targets.Secondly,the detector module is restructured using grouped convolution and parameter-sharing strategies to reduce model parameters while maintaining detection accuracy.Lastly,an attention guidance module combining large separable kernel attention(LSKA)and bottleneck attention module(BAM)is introduced and integrated into the spatial pyramid pooling(SPPF)module.LSKA expands the receptive field,improving the model's ability to capture local features of occluded targets.At the same time,BAM focuses on both spatial and channel attention,effectively filtering environmental noise and reducing the influence of environmental interference on model detection.Experimental results show that compared to the baseline algorithm,the improved algorithm increases accuracy by 0.2 percentage points,recall by 2.1 percentage points,and mAP@0.5 by 2.6 percentage points.The model size is only 2.0 M,and the FPS reaches 84.5 frames.

关 键 词：安全帽检测 YOLOv8n 特征融合 改进小目标层 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]