摘要
实时的吸烟行为监测对于保障工地安全具有重要的意义。研究针对隧道等低照明环境因光照强度低、光线分布混杂、点光源过曝光而导致的烟支检测精度低的问题,提出了一种基于改进YOLOv5s(You Only Look Once Version 5s)的吸烟行为检测模型。首先,设计一种图像增强方法,旨在限制点光源过曝光产生的局部高光,增强烟支特征细节,改善图像对比度。其次,引入卷积块注意力模块(Convolutional Block Attention Module,CBAM),使模型更加聚焦烟支目标区域的内容信息和位置信息。最后,改进多尺度检测头,增加适用于烟支的更小检测层,提升模型对烟支的检测能力。试验结果显示,研究提出的针对低照明环境的检测模型可将平均检测精度从91.8%提升至95.9%,相较于原模型和其他经典模型,检测效果得到了显著提升,表明了方法的有效性。
To overcome the difficulties associated with low light intensity,mixed light distribution,overexposure of point light sources,and the small size of cigarette targets in dimly lit environments like tunnels,we have developed an advanced smoking behavior detection model based on the YOLOv5s framework.This approach is tailored to substantially enhance detection accuracy,especially in challenging lighting conditions.Initially,we devised a sophisticated image enhancement technique focused on finely optimizing the luminance and contrast of the visuals.This method rectifies uneven lighting conditions,balances areas that are underexposed and overexposed,and notably mitigates the adverse effects of low-light settings,thereby enhancing detection precision.Secondly,we have incorporated the CBAM(Convolutional Block Attention Module)attention mechanism into each C3 module of the backbone network.This addition enables the model to concentrate more on the content and spatial information of the cigarette target area.Consequently,the model's capability to extract target features is enhanced,while interference from complex background environments is suppressed.As a result,the impact of the environment on the accuracy of smoking detection is diminished.Finally,we have augmented the multi-scale detection head of the original model by introducing an additional smaller detection layer specifically designed for cigarette detection.This enhancement,including the existing large,medium,and small detection layers,bolsters the model's capability to detect cigarette targets at medium to long distances.Consequently,the overall detection accuracy is significantly improved.Through rigorous experimental validation,the enhanced model showcases superior precision,recall,and average precision when compared to other established models.In comparison to the original YOLOv5s base model,the improved model exhibits a 4.0 percentage point increase in precision(P),a 3.3 percentage point increase in recall(R),and a 4.1 percentage point increase in average precision,ultimately achieving an impressive rate of 95.9%.The enhanced model significantly strengthens the detection capability for smoking behavior in complex environments characterized by low lighting conditions,thus affirming the effectiveness of this approach.
作者
谭梦颖
王祥
聂磊
TAN Mengying;WANG Xiang;NIE Lei(School of Mechanical Engineering,Hubei University of Technology,Wuhan 430068,China)
出处
《安全与环境学报》
CAS
CSCD
北大核心
2024年第9期3522-3531,共10页
Journal of Safety and Environment
基金
湖北省科技创新人才计划项目(2023DJC048)
湖北省教育厅科技项目青年项目(Q20284409)
湖北工业大学博士科研启动基金项目(BSQD2020006)
湖北省自然科学基金项目(2022CFB473)。
关键词
安全工程
低照度图像
目标检测
图像增强
注意力机制
safety engineering
low-light image
object detection
image enhancement
attention mechanism
