基于人类视觉皮层双通道模型的驾驶员眼动行为识别  

Recognition of driver’s eye movement based on the human visual cortex two-stream model

作　　者：申天啸 韩怡园 韩冰[1] 高新波 SHEN Tianxiao;HAN Yiyuan;HAN Bing;GAO Xinbo(School of Electronic Engineering,Xidian University,Xi’an 710071,China;Chongqing Key Laboratory of Image Cognition,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]西安电子科技大学电子工程学院,陕西西安710071 [2]重庆邮电大学重庆市图像认知重点实验室,重庆400065

出　　处：《智能系统学报》2022年第1期41-49,共9页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61572384,62076190,41831072);西安电子科技大学研究生创新基金项目.

摘　　要：驾驶员的危险行为会增加交通事故的发生率,目前对驾驶员行为的研究中大多通过面部识别等方法对异常行为如疲劳驾驶、接电话等进行识别。这种方法仅客观地对驾驶员行为进行分类,而忽略了他们在驾驶过程中的主观心理。眼动仪是记录和分析驾驶员眼动数据的有效工具,可以清晰地了解驾驶员的想法并总结其视觉认知模式。因为目前还没有针对驾驶员眼动行为的数据库,首先构建了真实道路场景下的眼动视频数据集VIPDAR_5,与传统数据相比,它存在更多的摄像机运动、光照变化、视线遮挡等情况。针对这些问题提出了一个基于人类视觉皮层双通路的模型TWNet,通过模拟视觉机制,提高了驾驶员眼动行为的识别性能。另一方面,通过自适应最大池化层和通道权重设置,减少参数,提高准确率。在VIPDAR_5数据集上的实验结果表明,与现有方法相比,该模型能有效识别驾驶员眼动行为。Drivers’dangerous actions will increase the incidence of traffic accidents.The current researches on driver’s action are based on facial recognition to recognize abnormal actions,such as fatigue driving,cell phone usage.These methods only classify drivers’actions objectively and ignore their subjective thoughts during driving.The eye tracker is a device that can record and analyze driver’s eye movement effectively,understand their thoughts clearly and summarize their visual cognition patterns.There is no dataset for driver’s eye movement currently.Therefore,this paper first builds a eye movement video dataset named VIPDAR_5 applicable in real road scenes.Compared with traditional dataset,it contains more camera motion,illumination change,and sight occlusion situations.Therefore,the TWNet model based on two channels of the human visual cortex is built in this paper,which can improve recognition performance by simulating human visual mechanisms.On the other hand,adaptive max-pooling layer and channel weight setting are added to reduce parameters and improve recognition accuracy.Experimental results on the VIPDAR_5 dataset indicate that the model proposed in this paper can effectively recognize drivers’eye movement in comparison with existing methods.

关 键 词：眼动视频数据库 行为识别 深度学习 道路安全 辅助驾驶 眼动追踪 人类视觉系统 行为研究 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]