Gaze Estimation Algorithm Based On Appearance And Head Pose

With the widespread concern for the living conditions of the patients with “ALS(Amyotrophic Lateral Sclerosis)” and " Lou Gehrig's disease",the research of life support equipment has become one of the public concerns,too.As the conventional method of HCI(human-computer interaction)can't catch up with the needs of patients,The gaze-controlled method has become the first choice instead of it.The core of this method is the research and application of the gaze estimation technology.The technology of gaze estimation is the process of calculating gaze and location of gaze.As one of the most efficient and non-contact HCI method,it has great potential in every field.At present,the appearance-based methods of gaze-estimation has gradually become the key breakthrough in this field,yet it still faces the problem such as lack of algorithm and low accuracy.In this paper,a system of image acquisition and gaze-estimation is designed.The image is collected by the camera with a computer.After the preprocessing,the improved CNN(Convolutional Neural Network)is designed to train the gaze-estimation model,and the finally output the position of gaze point.The main job of this paper is as follows:1.An improved AdaBoost face detection algorithm.In order to improve the time efficiency while keeping the accuracy of the face detection.This paper introduces YCbCr color model for face detection ahead of AdaBoost cascade detection algorithm.With the help of the stability of the color model,the regions of faces will be detected first.This will be a good help to reduce the time of feature scanning and the real-time performance of detection is improved.The results of experiment shows that the time efficiency of the optimized face detection is obviously improved.2.Head pose estimation based on facial feature points.This paper opted the CLM to obtain face feature points,and then uses a flexible model to calculate the head posture.The CLM construction is divided into two parts: modeling progress and searching progress.Modeling progress needs to build the Shape-Model with the feature points distribution of known images,trained by PCA(Principle Component Analysis),and the Patch Model is trained by the SVM(Support Vector Machine)according to texture features in the images.Searching process optimized the response strategy of the response map for the constrained result nearest to the optimum solution.The method based on the flexible model is to estimate the head posture by comparing the location of the feature point of a plane image with the plane projection of 3D model.3.The gaze estimation algorithm based on shallow residual network.In order to improve the performance algorithm,this paper makes use of the residual block to deepen the network in order to improve the ability to extract potential features,and reducing the difficulty of training.The experiments of this paper separately trained models of traditional Le Net-5 and shallow residual network,with images from MPIIGaze dataset and the local dataset collected by the process mentioned in this paper.Comparing on the MPIIGaze and the local dataset,the accuracy is improved by 11.4% and 15%.The effect of improvement is proved.In conclusion,this paper has made a study of the algorithm of gaze-estimate and the related work.And on the basis of a simple hardware environment,build the gaze-estimation algorithm of CNN,improved algorithm with shallow residual network at the same time.The results of experiments with the public dataset and the local dataset has proved the improvement.