多角度视频的驾驶人员行为识别  

Research on Driver Action Recognition Based on Multi-Angle Videos

作　　者：赵维[1] 沈柏杉 张宇 孔俊 ZHAO Wei;SHEN Baishan;ZHANG Yu;KONG Jun(Department of Information Engineering,Jilin Police College,Changchun 130117,China;College of Information Science and Technology,Northeast Normal University,Changchun 130117,China)

机构地区：[1]吉林警察学院信息工程系,长春130117 [2]东北师范大学信息科学与技术学院,长春130117

出　　处：《吉林大学学报（信息科学版）》2020年第3期353-359,共7页Journal of Jilin University（Information Science Edition）

基　　金：国家自然科学基金资助项目(61672150);吉林省科技厅基金资助项目(20180201089GX,20190201305JC);吉林省教育厅基金资助项目(JJKH20190291KJ,JJKH20190294KJ,JJKH20190355KJ)。

摘　　要：驾驶员在行驶过程中看手机,与乘车人员交谈等违规行为,为安全行驶造成了极大的隐患。为了解决此类问题,提出了一种多角度行为识别方法,从3个角度同步捕捉驾驶员行为的视频,构建多角度驾驶员行为的视频和数据集,利用深度卷积神经网络,进行识别分类。实验结果表明,3D CNN相对于2D CNN的识别精度更加准确,在对比输入剪辑的帧数实验中,发现堆叠的视频帧数会影响准确度,并在具有较大优势的R2plus1D模型中(将3D卷积滤波器分解为单独的空间和时间分量),基于多角度驾驶人员行为识别精度达到87%。In the process of driving,violations such as looking at mobile phones and talking with passengers cause great hidden dangers to safe driving.In order to solve such problems,we propose a multi-angle behavior recognition method,which can synchronously capture video of driving behaviors from three angles,construct a multi-angle driver behavior data set,and use deep convolutional neural network to carry out recognition and classification.The experimental results show the advantages of 3D CNN over 2D CNN in accuracy.And in comparing input clip frames experiment,we found the stacked video frames will affect accuracy and in R2plus1D model(3D convolution filter is decomposed into a separate component of time and space),realized driver behavior based recognition accuracy of 87%from multiple perspectives.

关 键 词：机器视觉 深度学习 行为识别 驾驶人员 多角度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]