轻量化的驾驶员吸烟检测模型  

Lightweight driver smoking detection model

作　　者：陈新一 王晓强[1] 李少波 夏旭 陶乙豪 庄旭菲[1] CHEN Xin-yi;WANG Xiao-qiang;LI Shao-bo;XIA Xu;TAO Yi-hao;ZHUANG Xu-fei(College of Information Engineering,Inner Mongolia University of Technology,Hohhot 010080,China)

机构地区：[1]内蒙古工业大学信息工程学院,内蒙古呼和浩特010080

出　　处：《计算机工程与设计》2025年第2期626-632,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62165011);内蒙古自然科学基金项目(2023MS06021);2023年自治区直属高校基本科研业务费基金项目(JY20230065);内蒙古自治区科技计划基金项目(2020GG0104)。

摘　　要：为进一步实现在移动设备上对驾驶员吸烟的违法行为实时检测,提出一种改进的SlimYOLOv7-tiny模型。使用DSConv替换YOLOv7-tiny特征融合网络中标准3×3卷积,增加P6检测层,采用Mish激活函数,引入边框回归损失函数EIoU loss,利用Slim剪枝算法进一步提高模型的轻量化,使用PyQt5开发图形界面程序。实验结果表明,模型在自建驾驶员吸烟数据集上与原模型相比参数量减少60.0%,计算量减小64.39%,有利于模型进一步在移动设备及嵌入式设备上的实时性检测。To further facilitate the real-time detection of illegal smoking behavior by drivers on mobile devices,an enhanced SlimYOLOv7-tiny model was proposed.DSConv was employed to replace the conventional 3×3 convolutions in the YOLOv7-tiny feature fusion network.A P6 detection layer was added,the Mish activation function was applied,and the EIoU loss for bounding box regression was introduced.The model lightweighting was achieved through the utilization of the Slim pruning algorithm.A graphical user interface program was developed using PyQt5.Experimental results indicate that,when evaluated on a custom dataset of smoking drivers,the model exhibits a 60.0%reduction in parameter count and a 64.39%decrease in computational load compared to the original model.This is conducive to the model’s enhanced real-time detection capabilities on mobile and embedded devices.

关 键 词：驾驶员吸烟行为 目标检测 实时 激活函数 轻量化 图形界面 剪枝算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]