Research On Recognition Of Action Units For Fatigue Expression Based On Optical Flow And HMM

Currently, expression recognition is become one of the hot research topics, and it is important significant for the analysis of emotion, human-computer interaction and intelligent system. Facial action unit identification is the basis of facial expression recognition. It provides more detailed analysis for the different facial characteristics and mental states. Facial expression contains a wealth of information, but most of the researchs on the AUs recognition focus on the traditional basic expressions. It has broad prospects for expanding expression recognition research and application field by studing in the specific expressions’ feature and recognition such as fatigue, confusion and so on.In this paper, we analyse the fatigue feature deeply and try to find the AUs for describing the fatigue face. The FACS decomposes different facial expression into independent action units, and using AUs coding to sign the action unit. It provides important reliable bases for sentiment analysis and expression modeling to analyse the action in fatigue facial expression and find the AUs which can describe the facial movements. But the FACS is just a heuristic information system, and AUs are purely static spatial models. Expressions and other facial movements are dynamic processes combined with different facial muscle. In this paper, we adopt image sequences to extract the features of the movements and identify dynamic facial AUs and highlight the dynamics effectively.Here we use optical flow algorithm to extract the dynamic characteristics. The optical flow holds the object’s movement information. It can be used for describing the movements of the target object. Before the processing we use the AMM method to locate facial feature points and, It can remove redundant information and reduce the amount of calculation effectively by feature regions divisions, so we can focuse on analysis of characteristics of human face. Optical flow feature datas have high dimension and not fit for the direct classification processing, thus using PCA method to reduce the dimension. By extracting the coefficient of PCA projection to achieve the purpose of dimension reduction, and construct the eigenvalue sequences by combining the PCA coefficient of multiple AU image frames. For the characteristic sequence observations are multidimensional row vector, continuous hidden markov model is used to establish the AU models. The observation probability of continuous hidden markov model is established by Gaussian mixture model. The states’ number of the model, and the Gaussian mixture coefficients obtained from the experiments.Experiments are tested on the feature value generation scheme. Also chose suitable parameters to establish the AU HMM. Finally, experiments test and analyze the recognition results for fatigue facial samples and normal facial samples.The experiments indicate that the algorithm this paper proposed can identify the AU and simple AU combinations in fatigue face, but has not so good identify results for the subtle movement of AUs.