基于轻量化网络与嵌入式的分心行为协同检测系统  

Collaborative Detection System for Distraction Behavior Based onLightweight Network and Embedded Platform

作　　者：李少凡 高尚兵 张莹莹[1,2] 黄想 杨苏强 郭筱宇 Li Shaofan;Gao Shangbing;Zhang Yingying;Huang Xiang;Yang Suqiang;Guo Xiaoyu(Faculty of Computer and Software Engineering,Huaiyin Institute of Technology,Huai’an 223001,China;Laboratory for Internet of Things and Mobile Internet Technology of Jiangsu Province,Huaiyin Institute of Technology,Huai’an 223001,China)

机构地区：[1]淮阴工学院计算机与软件工程学院,江苏淮安223001 [2]淮阴工学院江苏省物联网移动互联技术工程实验室,江苏淮安223001

出　　处：《南京师范大学学报（工程技术版）》2023年第1期25-32,共8页Journal of Nanjing Normal University(Engineering and Technology Edition)

基　　金：国家重点研发计划项目(2018YFB1004904);国家自然科学基金面上项目(62076107);江苏省高校自然科学研究重大项目(18KJA520001);江苏省产学研合作项目(BY2022334);淮阴工学院研究生科技创新计划项目(HGYK202216)。

摘　　要：分心驾驶是交通事故发生的主要原因之一.针对目前分心驾驶检测手段单一、检测种类少、检测效率低的问题,提出一种基于轻量化网络与嵌入式的分心行为协同检测系统.首先,结合Ghost模块和通道注意力机制提出一种轻量化目标检测网络YOLO-Ghost,采用CSPGBottleck构建GhostDarknet作为主干网络,同时构建一种具有多尺度注意力机制的多特征融合模块SE-FPN来进行特征融合,根据固有检测场景进行检测头优化,以CIOU(complete-IOU)作为损失函数.采用YOLO-Ghost识别和定位局部特征,提出APJ(anchor position judge)对手动分心行为进行判定;协同检测方面,利用MobileNetv3与YOLO-Ghost协同进行人脸关键点回归和视线估计;最后利用检测出的多模态信息对驾驶员当前行驶状态进行联合判定.实验结果表明,YOLO-Ghost的准确率和检测速度优于其他主流方法.将算法部署到嵌入式设备中,在NVIDIA Jetson TX1上实现了20FPS的实时检测性能,准确性和实时性均达到检测要求.Distracted driving is the main cause of traffic accident.In order to solve the problems of fewer kinds of distracted driving detection and poor detection efficiency,a collaborative detection system for distraction behavior based on the lightweight network and embedded platform is proposed.First of all,a lightweight object detection network YOLO-Ghost is proposed by combining Ghost module and channel attention mechanism,the CSPGBottleck is proposed to build GhostDarknet as the backbone network,and a multi-feature fusion module SE-FPN with a multi-scale attention mechanism is proposed for feature fusion.A more comprehensive CIOU(complete-IOU)function is considered as the loss function.YOLO-Ghost is used to identify and locate local features,and APJ(anchor position judge)is proposed to judge manual distraction behavior.Secondly,MobileNetv3 and YOLO-Ghost are used to perform face key point regression and gaze estimation.Finally,the detected multimodal information is used to jointly determine the current driving state of the driver.The experimental results show that the YOLO-Ghost achieves the higher accuracy and speed than other main stream methods.At the same time,when the algorithm is deployed to the embedded device,it obtains 20FPS real-time detection performance on the NVIDIA Jetson TX1 and the accuracy and real-time performance reach the detection requirements.

关 键 词：协同检测 人物交互 轻量级网络 智能交通 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]