基于YOLOv11n的现场作业人员违规行为检测算法

The Algorithm for Detecting Violations of on-site Workers"" Behavior Based on YOLO

摘要:

在电力行业中，现场作业环境复杂且危险因素众多，作业人员的违规行为如未佩戴安全帽和穿反光防护服等极易引发严重事故。传统的人工巡查和视频回放等安全监管方式存在主观性强、反应滞后和监控效率低等问题，为此，提出了一种基于改进 YOLOv11n 的安全帽与反光衣穿戴规范检测算法。该算法设计了多尺度边缘信息生成器模块，将浅层特征中的边缘细节信息与不同深层特征融合，增强了模型对边缘信息的提取能力；引入 BiFPN，通过可学习权重动态调整不同层级特征的贡献，更好地保留小目标特征；改进了检测头，在降低模型参数量的同时提升了准确度；使用 MPDIoU 替换原本的损失函数，有效增强了对不规则目标的检测能力并加速了收敛过程。实验结果表明，与基准模型YOLOv11n相比，改进模型在 mAP50 指标上表现更优，达到了94.5%，验证了其在现场作业人员违规行为检测中的有效性和优势。

Abstract:

In the power industry, the on-site operation environment is complex and full of risk factors. Violations of operating personnel, such as not wearing safety helmets and reflective protective clothing, are prone to cause serious accidents. Traditional safety supervision methods such as manual inspection and video playback have problems such as strong subjectivity, lagging response, and low monitoring efficiency. Therefore, this paper proposes a detection algorithm for safety helmet and reflective vest compliance based on improved YOLOv11n. The algorithm designs a multi-scale edge information generator module to fuse the edge detail information in the shallow features with different deep features, enhancing the model's ability to extract edge information. BiFPN is introduced to dynamically adjust the contribution of features at different levels through learnable weights, better preserving small target features. MPDIoU is used to replace the original loss function, effectively enhancing the detection ability for irregular targets and accelerating the convergence process. The detection head is improved to reduce the model's parameter count and improve accuracy. Experimental results show that the improved model outperforms the baseline model YOLOv11n on the mAP50 metric, achieving 94.5%, which validates its effectiveness and superiority in detecting violations among on-site operational personnel.