Reaserch On Face Analysis Method For Driver Fatigue Detection

Fatigue driving is one of the important causes of traffic accidents.Driver fatigue state estimation based on face video analysis has attracted extensive attention in fatigue detection research because of its non-invasive characteristics.Establishing fatigue evaluation indexes by quantitatively analyzing the state variation of face or components,such as eyes and mouth,is the most widely used and recognized objective method at present.However,due to the influence of partial occlusion,local noise,illumination variation,pose variation and other factors,it is difficult to obtain the correct and effective feature representation for whole face and local area in the actual driving environment,which greatly increases the difficulty of face state analysis.Aiming at the key problems such as face alignmnet and pose estimation under unconstrained conditions,judgment of eye motion and variation trend,temporal detection of blink behavior and so on,this paper study face analysis method for driver fatigue detection.The specific research contents and contributions are as follows:(1)A 2D face alignment method by integrating global matching and local refinement is proposed.In current face alignment methods,the local optimal matching results lack global constraints,and the global appearance is insufficiently represented in local details,which leads to large alignment error.Theref'ore,the face shape is respectively updated and weighted in different ways by encoding the multi-mode face information as the representation features of whole face or local regions,in which the global optimization and local detail correction are achieved jointly.The experimental results indicate that the proposed method can improve align accuracy under the condition of partial occlusion and local illumination variation.(2)The method of alignment and pose estimation for unconstrained face by using multi-level piecewise regression strategy is proposed.The traditional single regression method is difficult to accurately and efficiently build the high-dimensional nonlinear mapping from 2D images to 3D parameters,which leads to increasing errors or wasting computation.Therefore,according to the texture variation characteristics of different stages,the rigid parameters and elastic parameters reflect face variation are sequentially adjusted in different degree of fineness by applying gradually reduced regression scale.The experimental results indicate that the proposed method improves image description stability and enhances align accuracy and computing efficiency on a face with large pose deflection.(3)The method of quantitatively analysizing eye state and furtherly detecting blink by fusing multi-channel spatio-temporal information is proposed.Because the eye state judgment method based on single frame image ignores timing and motion information,it can not reflect the dynamic variation trend and lacks the overall constraints to the motion process,which leads to the decreased detection efficiency.Therefore,on the basis of intra-frame analysis,the different visual clues,such as motion information and spatio-temporal saliency information,are fused at the decision level for the results of eye state analysis from different channels.A more descriptive and discriminative quantitative representation is obtained for blinking detection.The experimental results indicate that compared with a single description feature,the proposed method improves accuracy and reliability of eye state quantitative analysis.(4)The method of blink behaviour detection in video sequence by using differential spatio-temporal multi-scale analysis is proposed.The blink motion has a non-uniform distribution in whole time series.The same processing method is used for the continuous eyes opening and the blink,which greatly increases the computing complexity.Therefore,blink is regarded as a specific behavior pattern and spatial-temporal multi-scale analysis is performed.According to the specific content of different segments,the differential feature representation and detection method are appled.The strategy of rejection layer by layer,discrimination and confirmation is performed to gradually reduce the detection scale and accurately locate blink process in the video.The experimental results indicate that compared with the traditional behaviour detection method,the proposed method reduces the computing complexity and achieves higher blink detection accuracy.The above research results are applied in a fatigue detection system based on video analysis,which can increases the precision of extracting geometric features,intra-frame image features and inter-frame timing features.Then the accuracy of eye state quantitative analysis and blink detection is enhanced.Finally,the effectiveness and reliability of fatigue detection are improved,which helps to application and popularization of practical system.