基于改进YOLO-v3的眼机交互模型研究及实现  

Research and Implementation of Eye Computer Interaction Model Based on Improved YOLO-v3

作　　者：陈亚晨 韩伟[1,2] 白雪剑 陈友华 赵俊奇[1,2] 阎洁 CHEN Ya-chen;HAN Wei;BAI Xue-jian;CHEN You-hua;ZHAO Jun-qi;YAN Jie(Information and Communication Engineering College,North University of China,Taiyuan 030051,China;Shanxi Optometry Biological Diagnosis and Treatment Equipment Engineering Research Center,Taiyuan 030051,China;Shanxi Provincial Key Laboratory of Biomedical Imaging and Imaging Big Data,Taiyuan 030051,China)

机构地区：[1]中北大学信息与通信工程学院,太原030051 [2]山西省视光学生物诊疗设备工程研究中心,太原030051 [3]生物医学成像与影像大数据山西省重点实验室,太原030051

出　　处：《科学技术与工程》2021年第3期1084-1090,共7页Science Technology and Engineering

基　　金：国家自然科学基金(61505179);国防科技创新特区项目。

摘　　要：针对嵌入式眼-机交互技术中所采用的传统眼行为识别方法准确率低、速度慢等问题,并结合所研制眼机交互系统硬件特点及应用场景,提出一种基于改进YOLO-v3的眼机交互模型。该模型通过去除13×13特征分辨率的检测模块、增加浅层网络的层数以及采用K-means聚类算法选取初始先验框,提高了网络像素特征提取细粒度并加快了检测速度,进而结合人眼特征参数提取方法和眼行为识别算法,构建出了眼机交互模型并进行实验。实验结果表明,该模型对不同眼行为的识别率达91.30%,改进的YOLO-v3网络的平均检测准确率(mean average precision,mAP)为99.9%,识别速度达22.8 FPS,相比原YOLO-v3方法检测时间缩短了11.4%。Aiming at the low accuracy and slow speed of the traditional eye behavior recognition methods used in the embedded eye computer interaction technology,combined with the hardware characteristics and application scenarios of the developed eye computer interaction system,an eye computer interaction model based on the improved YOLO-v3 was proposed.By removing the 13×13 feature resolution detection module,increasing the number of layers of shallow network and using K-means clustering algorithm to select the initial candidate box,the fine-grained feature extraction of network pixels was improved and the detection speed was sped up,and then combining the extraction method of human eye feature parameters and eye behavior recognition algorithm,an eye machine interaction model and experiments were built and carried out.The experimental results show that the recognition rate of the model is 91.30%,the mean average precision(mAP)of the improved YOLO-v3 network is 99.9%,and the recognition speed is 22.8 FPS.Compared with the original YOLO-v3 method,the detection time is shortened by 11.4%.

关 键 词：眼机交互 YOLO-v3 实时性 特征提取 眼行为识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]