Research On Multi-feature Fatigue Driving Detection Based On Improved MTCNN Network

With the improvement of the national standard and the development of the transportation industry,private cars are becoming more and more popular,accompanying it is the frequent occurrence of traffic accidents,which in turn has a big impact on safety of human life and property.Related reports show that fatigue driving is an important cause of traffic accidents.If we can detect the driver’s fatigue state in real time and give prompt reminders,can we reduce the traffic accident rate in a certain sense.Among current detection fatigue driving detection methods,the detection method based on the physiological characteristics of the driver needs to wear related detection equipment,which has a high cost and will affect the driving state.The detection method based on the behavior characteristics of the motor vehicle is vulnerable to weather,road conditions,driving environment and other factors,it has low robustness.Compared with the above two methods,the fatigue detection method based on facial features has the advantages of non-contact,non-interference and high detection accuracy.Therefore,this article focuses on this method for research.The content mainly includes the following aspects:(1)Face detection and feature point positioning.This paper uses the MTCNN algorithm based on the multi-task cascade framework in order to accomplish face detection and coarse positioning of five feature points.First,select the WIDER FACE and Celeb A data sets to train the MTCNN network.Experiments show that after MTCNN is trained by the three sub-networks P-Net,R-Net and O-Net,the accuracy rate has reached 97.3% from 95.8%.Secondly,this article optimizes the parameters of the three sub-networks on the basis of the original network: For the P-Net network,a convolutional layer with a 3×3 convolution kernel is added;For the R-Net network,the fully connected layer is replaced with mean pooling,and the convolutional kernel of the last layer is changed from 128 to 196;For the O-Net network,change the parameters of the first,second,and fourth convolutional layers,and set the size of the maximum pooling layer to 2×2.After optimization,the accuracy of the MTCNN network has increased from 97.3% to 98.92%,an increase of 1.62%.After obtaining the face candidate frame,this paper selects the 300-W face database to train the cascade regression algorithm ERT,finally,the 68 key points of the face are accurately positioned.(2)Face tracking.This paper designs a face tracking algorithm based on the combination of DSST and TLD,Because DSST lacks a detection mechanism for tracking failure cases,TLD is susceptible to external factors such as light.These two methods can complement each other,so this article merges two face tracking algorithms: Replace the tracking module in the TLD with DSST and a tracking failure detection method suitable for this algorithm,and add the HOG feature to the detection module of the TLD algorithm.Through experiments,various facial postures of the driver with naked eyes,glasses and partial occlusion are simulated,experiment with the MTCNN network and the face tracking algorithm of DSST+TLD.The results show that in these complex environments,the algorithm designed in this paper shows excellent performance.(3)Fatigue feature extraction.After obtaining the position of the facial feature points,analyze the three features of the driver’s eyes,mouth,and head posture.Aiming at the extraction of eye features,this article selects the eye aspect ratio EAR algorithm,and its value is calculated by averaging two eyes.The PERCLOS and blinking frequency are obtained by this algorithm,we use it as the basis for judging fatigue;For the extraction of mouth features,the MAR algorithm is used.In view of the different thickness of the lips of each person,this paper proposes to use the inner contour of the mouth to calculate the MAR value,and yawn as the basis for judging fatigue;For the head feature extraction,the pitch angle in the Euler angles in the three-dimensional space is selected to judge the fatigue state.Select 5 people to simulate the fatigue state,conduct experiments on the above fatigue characteristics,and obtain the fatigue threshold: P80 is 0.25,EAR is 0.23,MAR is 0.6,and the pitch offset is20%.(4)Multi-feature fatigue detection.Due to the low accuracy of most single fatigue feature extraction at present,this paper designs a multi-feature fatigue detection.First,the device with a resolution of 1080 p HD and a video frame rate of 30 fps is used to collect videos of the awake and fatigue states of 10 experimenters,intercept 5 fatigue and non-fatigue videos of two minutes each,randomly select 250 groups of samples to train the support vector machine,and test on the remaining data set,compare different detection algorithms under the same experimental conditions.Experimental analysis shows that the accuracy of the method used in this paper can reach 96.42%,which is more feasible for fatigue driving detection.Finally,use Python language and Tensorflow framework to program the algorithm of this article to complete the fatigue driving detection system.