基于改进Faster RCNN的驾驶员手持通话检测  

Hand-held call detection of driver based on improved Faster RCNN

作　　者：王彬[1,2] 李小曼 赵作鹏 WANG Bin;LI Xiaoman;ZHAO Zuopeng(School of Computer Science and Technology,China University of Mining and Technology,Xuzhou,Jiangsu 221116,China;Department of Information Technology,Jiangsu Union Technical Institute,Xuzhou,Jiangsu 221008,China)

机构地区：[1]中国矿业大学计算机科学与技术学院,江苏徐州221116 [2]江苏联合职业技术学院信息技术系,江苏徐州221008

出　　处：《江苏大学学报（自然科学版）》2023年第3期318-323,共6页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61976217);徐州市重点研发项目(KC18082)。

摘　　要：针对现有驾驶员通话行为识别误判率较高的问题,提出一种基于改进Faster RCNN的驾驶员行为检测方法,对驾驶员的违规手持通话进行检测.介绍了针对区域建议网络(RPN)及其损失函数的优化策略,并在原始Faster RCNN上运用多尺度训练、增加锚点数量以及引入残差扩张网络的方法增强网络检测不同尺寸目标的鲁棒性.基于车载平台上采集的驾驶员行为图像,对文中提出的方法进行仿真试验.结果表明:RPN和Faster RCNN通过交替优化共享特征提取网络部分,实现高效的目标检测,相较于原始Faster RCNN,检测精确度提高了3.8%,对环境的适应性更强.To solve the problem of high false positive rate of existing driver call behavior recognition,an improved Faster RCNN was proposed based on driver behavior detection method for detecting the illegal hand-held call of driver.An optimization strategy for the region proposal network(RPN)and the loss function was introduced,and the robustness of the network in detecting targets with different sizes was enhanced by applying multi-scale training,increasing the number of anchor points and introducing the residual expansion network on the original Faster RCNN.The simulation experiments of the proposed method were conducted with the images of driver behavior collected on an in-vehicle platform.The results show that compared with original Faster RCNN,RPN and Faster RCNN can realize efficient target detection by alternatively optimizing the shared feature extraction network part with 3.8%improvement in detection precision and better adaptation to the environment.

关 键 词：驾驶员危险行为 目标检测 分神驾驶 驾驶辅助 多尺度训练 残差扩张网络 Faster RCNN 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]