基于ILF-YOLOv3的人员在岗状态检测算法研究  

On-the-job Status Detection Algorithm Based on ILF-YOLOv3

作　　者：谢斌红[1] 栗宁君 陈立潮[1] 张英俊[1] XIE Bin-hong;LI Ning-jun;CHEN Li-chao;ZHANG Ying-jun(Department of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机与技术学院,太原030024

出　　处：《太原科技大学学报》2021年第6期441-448,455,共9页Journal of Taiyuan University of Science and Technology

基　　金：山西省重点计划研发项目(201803D121048);山西省重点计划研发项目(201803D121055);山西省重点研发计划重点项目(201703D111027);山西省科技重大专项项目(20141101001)。

摘　　要：为了解决员工在岗状态的实时监测和管理问题,提出了一种改进YOLOv3的目标检测算法,ILF-YOLOv3(Improve Loss and Feature-YOLOv3).首先,使用二分交叉熵损失函数和添加制衡权重参数的方式对YOLOv3算法的损失函数进行改进;然后,增加了模型多尺度特征检测模块的特征融合密度;最后,针对采样数据集单一性的问题,采用生成式对抗网络对其进行定向增强。实验结果表明,改进后的算法在自制的StaffSData-Strong数据集上mAP值提高了7.9%,召回率提高了14%.In order to solve the problem of real-time monitoring and management of on-the-job status of employees, an improved object detection algorithm for YOLOv3,ILF-yolov3(Improve Loss and Feature-yolov3)is proposed.Firstly, the loss function of YOLOv3 algorithm is improved using binary cross entropy loss function and adding balanced weight parameters.Then, the feature fusion density of multi-scale feature detection module is increased.Finally, in order to solve the problem of unitary sampling data set, the generative antagonism network is used to enhance it.The improved algorithm increased the mAP value by 7.9% and the recall rate by 14% on the self-made staffsdata-strong dataset.

关 键 词：目标检测 YOLOv3网络 交叉熵损失函数 多尺度特征融合 ILF-YOLOv3网络 在岗状态检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]