基于改进YOLOv5s的复杂施工现场吸烟检测

Smoking Detection at Construction Site Based on Improved YOLOv5s

在建筑工地等复杂施工现场吸烟可能会引起火灾、爆炸等事故,严重危害施工安全。为了实现对建筑工地等施工现场的吸烟检测,使用YOLOv5s对人脸和烟支进行检测,并根据人脸和烟支的位置关系判断施工现场是否存在吸烟行为。为了提高人脸和烟支的检测精度,该文在原始模型的基础上做了三个改进:一是采用SIMOTA动态标签分配方法,提高了网络的召回率和检测速度;二是引入了尺度内特征交互模块AIFI,增强了网络的特征表达能力;三是使用了动态卷积ODConv,优化了特征提取模块C3,提高了网络的精确率。在自制数据集上进行实验,改进后的网络在精确率、召回率和平均精度方面均提升2%以上,检测速度提升了22%,取得了明显的性能优势。与主流算法对比,改进后的算法在检测速度和网络性能上均有明显优势,满足了施工现场吸烟检测的需求。

Smoking at construction sites and other complex construction sites may cause fire,explosion and other accidents,seriously endangering construction safety.In order to achieve smoking detection at construction sites and other construction sites,we use YOLOv5s to detect faces and cigarettes,and judge whether there is smoking behavior at the construction site according to the position relationship between faces and cigarettes.In order to improve the detection accuracy of faces and cigarettes,we make three improvements on the basis of the original model.First,the dynamic label assignment method,SIMOTA,is adopted,which improves the network recall rate and detection speed.Second,the scale-in feature interaction module,AIFI,is introduced,which enhances the network feature expression ability.Third,dynamic convolution,ODConv,is used to optimize the C3 feature extraction module,which improves the network accuracy.Experiments on self-made data sets show that the improved network has improved by more than 2%in accuracy,recall rate and average precision,and the detection speed has increased by 22%,achieving obvious performance advantages.Compared with the mainstream algorithms,the improved algorithm has obvious advantages in detection speed and network performance,meeting the needs of smoking detection at construction sites.