湖北汽车工业学院学报JournalofHubeiUniversityofAutomotiveTechnology第37卷第1期2023年3月Vol.37No.1Mar.2023doi：10.3969/j.issn.1008-5483.2023.01.008基于多特征融合的疲劳驾驶检测王康，樊继东（湖北汽车工业学院汽车工程学院，湖北十堰442002）摘要：针对疲劳驾驶检测的特征源单一、辨识率低和实时性差等问题，提出基于多特征融合的疲劳驾驶检测方法。通过SSD目标检测算法进行人脸检测，利用轻量级模型PFLD实现人脸关键点定位。以眼部纵横比、嘴部纵横比和头部姿态为疲劳特征源，提取相关特征，对不同驾驶员疲劳阈值进行标定，基于改进的PERCLOS算法实现疲劳驾驶判定。仿真结果表明：多特征融合疲劳检测系统对自建数据集和YAW数据集的疲劳特征辨识率分别达到了90.5%和94.12%，在实时视频流上的执行效率达到31.59ms，实现疲劳预警。关键词：疲劳驾驶；人脸关键点检测；PERCLOS中图分类号：TP391；U471.15文献标识码：A文章编号：1008-5483（2023）01-0039-06FatigueDrivingDetectionBasedonFacialMultipleFeatureFusionWangKang,FanJidong（SchoolofAutomotiveEngineering,HubeiUniversityofAutomotiveTechnology,Shiyan442002,China）Abstract：Aimingattheproblemsoffatiguedrivingdetectionmethodssuchassinglefatiguecharacter⁃istics,lowfatiguerecognitionrateandpoorreal-timeperformanceinfatiguedrivingdetection,afatiguedrivingdetectionmethodbasedonmulti-featurefusionwasproposed.TheSSDtargetdetectionalgo⁃rithmwasusedforfacialdetection,andthelightweightmodelPFLDwasusedtorealizethelocationoffacialkeypoints.Theeyeaspectratio,mouthaspectratioandheadposturewereusedasfatiguefeaturesources,andfatigue-relatedfeatureswereextracted.Aimingatthedifferencesinfacialfeatures,thefa⁃tiguethresholdcalibrationofdifferentdriverswascarriedout.BasedontheimprovedPERCLOSalgo⁃rithm,thefatiguedrivingjudgmentwasrealized.Thesimulationresultsshowthat,thefatiguefeaturerecognitionrateofthemulti-featurefusionfatiguetestingsystemontheself-builtdatasetandtheYAWdatasetreaches90.5%and94.12%respectively.Theexecutionefficiencyonthereal-timevideostreamreaches31.59ms.Andthefatiguewarningissuccessfullyrealized.Keywords：fatiguedriving;facialkeypointdetection;PERCLOS收稿日期：2022-03-22...
