Research On Algorithms Of Face Anti-spoofing And Gaze Estimation In Automobile Intelligent Cockpit

The automotive intelligent cockpit system based on deep learning aims to use computer vision technology to build a more intelligent and safe driving experience.Among them,face anti-spoofing technology and gaze estimation technology are two important directions in the system.However,from the current research,the face anti-spoofing and driver’s gaze estimation are not yet mature in the application of smart cockpits.The reason is that,on the one hand,ordinary RGB cameras cannot capture the depth information of human faces and modeling of single modal data is difficult to achieve the desired effect.The degeneration leads to the low robustness of the existing live detection model in the actual environment,which brings security risks to face recognition.On the other hand,the driver’s gaze estimation technology cannot fully explore gaze-related information using only eye image modeling,and the human eye animation characteristics of the existing synthetic data sets,resulting in limited accuracy of the existing gaze estimation methods based on deep learning.In view of the above problems,researches are carried out on two aspects of face anti-spoofing and gaze estimation in the intelligent cockpit system.The main research work and innovations are as follows:(1)Research on the face anti-spoofing scheme based on TOF camera.The TOF depth camera is used to form a large-scale data set.The data set includes deep face images and infrared face images.Then the lightweight network MobileNet is used to learn and model these two kinds of data,which fully explores the difference between living and non-living.The essential difference between depth information and infrared texture information is finally combined with the information of these two different modalities to improve the performance of the model.(2)Research on the gaze estimation algorithm based on multi-region feature fusion.First,a monocular image selection algorithm based on head posture is proposed to obtain higher-quality eye images.Second,a fusion network is designed to jointly learn gaze information from eye and face images.The eye feature extraction network introduces a residual structure,and the facial feature extraction network introduces a spatial weights mechanism on the basis of AlexNet,which further improves the feature extraction ability of the network.Experiments on public data sets show that the proposed algorithm achieves a lower average error rate.(3)Finally,an improved model-based gaze estimation scheme is studied.First,aiming at the problem that the eye images of the synthetic data set are not real enough,which leads to the poor performance of the trained model in the actual environment,data enhancement methods are used to make the synthetic data closer to real human eye images.Secondly,through rational use of hourglass blocks and improved up-sampling method,a heat map regression network of eye key points is designed,and finally the gaze direction is calculated by key points.Experiments show that the model not only has high accuracy on the public data set,but also has high robustness and operating speed when tested in an actual environment.