Research On Driver Fatigue Detection Based On Hybrid Measures

With the enhancement of economy and the rapid increase of vehicle conservation,traffic accidents, especially the serious traffic accidents occurred more and more frequently.Researches show that most of highway accidents are closely related to driver fatigue.Therefore, researches on real time driver fatigue detection system have great significance toimprove road traffic safety.The fusion of driving performance and eye movements overcome the limitation ofsingle measure approach, analysis the correlation and complementarities. At the same time,the fusion measure has proved to be the most promising technology due to the non intrusive,good accuracy and real time performance. However, the development of driver fatiguedetection system has been seriously hindered for the unapparent driving performance, thevariability of driving scenes and individual differences of drivers. This paper focuses on keyissues in analysis of fatigue driving characteristics, extraction characteristics parameters,selection best time window, information fusion and development driver fatigue model. Areal time driver fatigue detection model was developed and tested. The specific researchcontents are as follows:1. Experiment program designation and data collection. Driving performance data andeye movements data from20drivers under different driving state were collectedsynchronized based on driving simulation. Fatigue level was scored on the KSS standard bythree researchers, driving performance data and eye movements data were coded andcatalogued according to the KSS score, training and test database for different fatigue levelwas established.2. Driving performance characteristic parameters extraction. According to a thoroughanalysis on the fluctuation characteristics of the steering wheel angle, steering wheel angularvelocity, throttle position and vehicle velocity when a driver becomes fatigue, quantitativecharacteristic parameters for detecting diver fatigue were extracted from time sequenceinformation of these parameters. The differences level of each characteristic parameters weretested by the analysis of variance (ANOVA), with varying fatigue levels. The best timewindow for each parameter was selected. The extraction driving performance characteristicparameters include: SAMEAN, SASTD, SAQ1MEAN, SAQ3MEAN, SE, SAVMEAN, SAVSTD and PNS.3. Eye movements characteristic parameters extraction. According to a thoroughanalysis on the change law of eye close behavior, gaze behavior and glance behavior when adriver becomes fatigue, eye movements characteristic parameters were extracted. Thedifferences level of each characteristic parameters were tested by the analysis of variance(ANOVA), with varying fatigue levels. The best time window for each parameter wasselected. The extraction eye movements characteristic parameters include: P80, AECS,MECD, BF and GTFNRD.4. General detection model development. In order to get the best subset from theuniverse of the measures, an optimized measures selection algorithm was established. Thisalgorithm took the performance of support vector machine algorithm (SVM) as evaluationcriterion and used the search strategy of sequential forward floating selection algorithm(SFFS) to select the optimal measures combination from the driving performance measuresand the eye movements measures. General detection model was developed based on theoptimal characteristic parameters. During the fusion process, slip time window was used toconfuse the characteristic parameters with different best time window, real time performancewas improved in this method. General detection model had a good performance, the averageaccuracy was82.27%, sensitivity was82.54%and specificity was81.94%.5. Adaptive detection model development. The impact of individual differences ondriver fatigue detection was qualified based on paired samples T test and ANOVA. Due toself stability, reference measure was extracted using alert driving data, and then individualparameters were calculated. Adaptive detection model was developed based on individualparameters. During the initial stage of driving, general model was used to detect driver’sstate and adaptive detection model was developed, after that, adaptive detection was used todetect driving state. This model detected driver fatigue state reaches an accuracy of88.15%,a sensitivity of88.25%and a specificity of88.02%.Key issues in driver fatigue detection were studied deeply in this paper. The impacts offatigue driving on driving performance and eye movements were analyzed, and characteristicparameters were extracted. General model and adaptive model were developed for driverfatigue detection in real time. The research results provide theoretical and technical supportto driver fatigue detection systems, and improve traffic safety.