改进YOLOv4-Tiny的面向售货柜损害行为人体检测  

Human Detection of Damage Behavior for Vending Cabinets Based on Improved YOLOv4-Tiny

作　　者：殷民 贾新春[1] 张学立 冯江涛[1] 范晓宇 YIN Min;JIA Xinchun;ZHANG Xueli;FENG Jiangtao;FAN Xiaoyu(College of Automation&Software Engineering,Shanxi University,Taiyuan 030013,China)

机构地区：[1]山西大学自动化与软件学院,太原030013

出　　处：《计算机工程与应用》2024年第8期234-241,共8页Computer Engineering and Applications

基　　金：国家自然科学基金(U1610116,61973201)。

摘　　要：无人货柜的安全检测一直是零售领域的热点话题。针对现有人工监控无法及时且有效地捕捉到部分消费者对自助售货柜及其内部商品的损坏行为这一问题,提出了一种改进YOLOv4-Tiny的面向售货柜损害行为人体检测方法。将真实场景采集到的监控视频进行预处理,完成对数据集DMGE-Act的制作,解决场景图像数据源不足的问题。提出了基于YOLOv4-Tiny的改进模型——YOLOv4-TinyX,通过修改神经网络的激活函数进行平滑逼近,分别在主干特征提取网络的最大特征提取层后引入CBAM,在加强特征提取网络中的上采样操作层后引入CA两种不同的注意力机制模块,并且进行了数据不平衡的修正,有效提升了算法的特征提取与检测能力。通过对比实验分析,改进后的模型参数量仅增加2×10^(4)的同时,平均精度均值mAP提升了10.29个百分点,结果表明该算法保持轻量化且对损害行为的检测精度有显著提升。The safety inspection of unmanned containers has always been a hot topic in the retail field.Aiming at the prob-lem that the existing manual monitoring cannot timely and effectively capture the damage behavior of some consumers to the self-service vending cabinet and its internal products,an improved YOLOv4-Tiny oriented human detection method of damage behavior for vending cabinets is proposed.First of all,the surveillance video collected in the real scene is prepro-cessed,and the production of the data set DMGE-Act is completed to solve the problem of insufficient scene image data sources.Then,an improved model based on YOLOv4-Tiny,YOLOv4-TinyX,is proposed.By modifying the activation function of the neural network for smooth approximation,CBAM is introduced after the largest feature extraction layer of the backbone feature extraction network,and in strengthening the feature extraction network.After the upsampling opera-tion layer,two different attention mechanism modules of CA are introduced,and the data imbalance is corrected,which effectively improves the feature extraction and detection capabilities of the algorithm.Through comparative experimental analysis,the improved model parameters are only increased by 2×10^(4),while the average precision mAP is increased by 10.29 percentage points.The results show that the algorithm remains lightweight and the detection accuracy of damage behavior is significantly improved.

关 键 词：无人值守 损害行为 YOLOv4-Tiny 平滑逼近 注意力机制 轻量化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]