Research On Face Feature Detection And Fatigue State Recognition

Fatigue driving detection based on machine vision has very important theoretical and practical value, and is closely related with people’s lives and property safety. Over the past few years, fatigue driving detection technology based on machine vision has made substantial progress, and a large number method related fatigue driving detection have been reported. However, the existing driver fatigue detection systems still have many technical problems needed to be resolved. In this paper, facial features localization and fatigue analysis in driver fatigue detection are mainly analyzed and researched. The main contents and innovations of this paper are as follows:(1) Proposed a face detection method across multi-poseTo overcome the problem that the traditional Gaussian skin model is not very robust in skin segmentation in different skin colors and different illuminations, an improved Gaussian skin model is proposed. The parameters of this proposed model can adapt to different faces. And this model combines Gaussian skin model and gray level distribution through weighting multiplication. Then, a new face detection method based on skin color model and Adaboost is proposed to detect faces. Experiments on FERET, LFW, GTFD and dataset of scene images including many faces, demonstrate that the proposed face detection method can adapt to multi-pose faces.(2) Improve Camshift to a rapid and robust face tracking methodThe Camshift tracking algorithm has good real-time, but the robustness of the detection process is not strong, while it is non-automatic. And the track may miss, when there is a shelter, another face or other interferences. The face detection of Adaboost algorithm and the prediction of Kalman Filering can boost the tracking of Camshift. The aim is to combine both algorithms efficiently to have high speed and high accuracy, and it can overcome different gestures tracking, face block, the presence of interference face and other issues. The Camshift algorithm is added with a feedback of Adaboost detection, which have great results in face tracking system.(3) Propose a rapid and accurate eye detection and pupil location methodIn the process of eye detection, in order to overcome the factors of expression changes, illumination changes and glasses’ shelters, the eye detection method based on Gabor filters and K-medoid algorithm is proposed. Initially the proposed method highlights the eyes’position with different scale Gabor filters. It then combines Gabor filter method and improved K-medoid algorithm to design cluster analysis to detect eyes’position. The lastly pupil locating is accomplished on the eye windows based on the method which combines gray level distribution and entropy function. Experiments with BioID database and FERET color database are used to evaluate this method. The experimental results demonstrate the consistent robustness and efficiency of the proposed method.(4) Proposed an efficient method to3D face pose estimationBased on analysis of the pro-existing face pose estimation methods, a new3D face pose estimation method based on Active Appearance Model(AAM) and T-Structure is proposed. Firstly, a set of AAM models can be obtained after training different poses faces. Then, the objective face is matched with the set of AAM models, to choose the optimum model to accurate position the face characteristic points. Finally, the T structure is built with the two eyes and the mouth, which is used to estimate the face pose. The experiments based on the Orient face dataset show that the method can adapt to large rotation angles, and can reach a good accuracy of3D face pose estimation.(5) Proposed a face parameter extraction method in fatigue driving detectionBased on the analysis of pre-existing methods to detect driver state, a new method to calculate parameters of the driver state based on Active Appearance Model (AAM) and3D face pose estimation is proposed, which is used in fatigue driver detection in this paper as a way to calculate parameters. Aim at overcoming the difficulty of low accuracy to apply AAM to different faces and poses, a new method of training AAM templates with one special person of different postures is proposed, which has higher accuracy. With the other difficulty of estimating3D face condition from2D AAM templates, the estimating method of combining AAM and3D face model is proposed. And the face state model is established to detect the state of eyes, mouth and blink. And the experiments on a talking video of Manchester University, a driving video in RS-DMV Dataset and an imitated fatigue driving video of our lab demonstrate the effectiveness of this method in estimating the driver’s state.