Research And Application Of Multi-pose Eye Location Algorithm Based On Deep Learning

Eye as one of the most prominent features of the human face,it can reflect the state of the person,the direction of sight and so on,the eye location is the prerequisite for obtaining these information.However,in the case of large changes in the head pose,the shape and contour of the eyes are different,which leads to the decrease of the accuracy of the eye location.Therefore,this paper focuses on the multi-pose eye location algorithm and applies it to fatigue driving detection.The main contents are as follows:(1)Multi-pose eye location algorithm based on improved mtcnn network.For the problem of the declining accuracy of eye location in multi-pose situations,this paper improves the structure of the Multi-task Convolutional Network(mtcnn)by cascades an eye regression network.Firstly,the image input into the mtcnn network model,and after the layer-by-layer screening,the position of the human face and the five key points(two eye centers,two mouth corners,and nose)can be obtained.Then,according to the position of eye center crop the local image of the eye and input into the regression network model to refine,after the network regression,a more accurate center position of the human eye can be obtained.(2)Application of eye location in fatigue driving detection.In order to apply the multipose eye location algorithm proposed in this paper to the actual fatigue driving detection,this paper a proposed fast eye ROI extraction method based on rectangle frame estimation.Firstly,calculating the ratio of the eye rectangle frame to the face rectangle frame by the statistically analysis of a large number of training data.Then,for each human face located at the center of eye,the eye rectangle frame could be calculated according to the relationship between the eye rectangle frame and the face rectangle frame.Finally,eye ROI images can input into eye status recognition to determine whether the driver is tired.The proposed multi-pose eye location algorithm based on improved mtcnn network achieves accuracy of 89.54% on the recognized challenging 300 w test set and 90.15% on largepose change CAS-PEAL data set.At the same time,the average positioning error is lower than the another method.The eye location algorithm proposed in this paper is applied to the fatigue driving system for video testing.Compared with the system using the traditional eye location method,the recognition rate of the system is increased from 87.8% to 91.3%.The experimental results show that the eye location algorithm proposed in this paper can improve the accuracy of eye location under multi-pose situations,and the fast eye ROI extraction method based on rectangle frame estimation can be well applied in actual fatigue driving detection.