摘要
随着无人超市中智能冰柜的需求量逐渐增大,智能冰柜商品识别技术得到迅猛发展,如无线射频识别、条形码识别、二维码识别及图像识别等。但由于信号干扰、条形码遮挡、二维码褶皱、识别网络模型过大等原因,以上识别技术存在辨识率不高,速度较慢的问题。针对以上问题,搭建激光散斑测距系统并提出一种基于YOLOv4改进算法的智能冰柜商品数量检测方法。该方法优化Backbone阶段CSP Darknet53中的卷积层和CSP模块,剔除Neck阶段8倍降采样,解决了YOLOv4算法模型体积较大、识别速度较慢的问题。通过IR相机获得激光散斑下红外图像,并运算得到深度图,进而与彩色图像配准得到商品深度信息,最终计算出商品剩余数量。该改进算法与YOLOv4、YOLOv3算法进行实验比较,结果表明改进后的YOLOv4模型mAP可达97.02%,模型大小仅为13.8 MB,相较于YOLOv4原有算法检测速度提升了50%。该方法可为商品库存远程管理提供一种高效的技术手段。
With the demand for intelligent freezers in unmanned supermarkets is gradually increasing,and the merchandise recognition technology of intelligent freezers is developing rapidly,such as radio frequency recognition,barcode recognition,two-dimensional code recognition and image recognition.However,due to signal interference,barcode obscuration,2D code folds and oversized recognition network model,the above recognition technologies have the problems of low recognition rate and slow speed.To address the above problems,a laser scattering distance measurement system is built and a method for detecting the number of goods in smart freezers is proposed based on the YOLOv4 improved algorithm.In the method,the convolutional layer and CSP module in the Backbone stage CSPDarknet53 are optimized,the 8-fold downsampling in the Neck stage us eliminated,and the problem of large model size and slow recognition speed of YOLOv4 algorithm are solved.The IR image under the laser scattering is obtained by using the IR camera,the depth map is obtained,and then the depth information is obtained with the color image,and the remaining quantity of goods is finally calculated.The improved algorithm is compared with YOLOv4 and YOLOv3 algorithms,and the results show that the improved YOLOv4 model mAP can reach 97.02%,and the model size is only 13.8 MB,which is 50% higher than the original YOLOv4 algorithm.This method can provide an efficient technical means for the remote management of commodity inventory.
作者
张斐
文灏淳
林浩彬
吴鹏
陈宝宇
项振坡
Zhang Fei;Wen Haochun;Lin Haobin;Wu Peng;Chen Baoyu;Xiang Zhenpo(School of Mechanical Engineering,Dongguan University of Technology,Dongguan,Guangdong 523808,China)
出处
《机电工程技术》
2024年第4期121-124,151,共5页
Mechanical & Electrical Engineering Technology
