Research On Driver Fatigue Detection Technology Based On Visual Feature Fusion

With the development of automobiles,the road traffic safety in China is becoming more and more important.Fatigue driving detection shows great practical value as the traffic accidents caused by fatigue driving increases fast.In this paper,fatigue driving detection based on computer vision is studied,and a fatigue detection system with high accuracy and robustness is designed.The main research contents are as follows.Firstly,driver's face is detected and tracked.After comparing and analyzing different kinds of algorithms,AdaBoost algorithm with low speed failing in the actual scene of fatigue driving detection is chosen to detect face in this paper.In order to accelerate face detection,the KCF algorithm with high accuracy and speed is used on the detected face area.In this paper,AdaBoost and KCF cascaded algorithms with speed at 44 FPS are used for face detection,showing the best compromise between real-time and high accuracySecondly,driver's facial feature points are located.After studying and analyzing various methods of facial feature point location,the SDM algorithm can obtain an accurate result even with facial occlusion or great change on facial posture and is chosen.Based on LFW-A&C data set,the correct positioning proportion of the SDM algorithm is 98% and suitable for fatigue driving detection system with the RMS error threshold at 6.After that,facial and heart rate features are extracted to characterize fatigue.Facial fatigue features are usually extracted from eyes and mouth.The open and close of the eyes and mouth are judged according to the position of facial feature points.The state changes of the eyes and mouth in a unit time are analyzed statistically to obtain characteristic parameters,i.e.blink frequency,PERCLOS value,the longest duration of closing eyes and yawning.In this paper,visual technology is used to extract the driver's real-time heart rate.The main steps include ROI decision-making and signal extraction based on the location of feature points,trend analysis based on L2,denoising by time-domain filtering and real-time heart rate acquisition by fast Fourier transform.And then the real-time heart rate changes of drivers are compared to obtain the characteristics of heart rate fatigue.The heart rate measurement experiment shows a good accuracy of this method at 94.67%.Finally,the above features are fused to detect fatigue state.In order to simplify the weight setting,three eye features are fused by linear weighting method,avoiding the excessive weight caused by the variety of eye features.Then,the eye,mouth and heart rate features are fused to detect fatigue.Compared with various single feature methods,the detection accuracy of this method is 87.5%,over 10% higher than that of others.The false detection rate and the missed detection rate have also decreased significantly to 7.0% and 5.5%,respectively.In this paper,KCF tracking algorithm is used to improve the real-time performance.In order to improve the detection accuracy,heart rate characteristic is introduced and feature fusion is carried out.To summarize,our method is applicable for the driver fatigue detection.