Facial Motion Analysis And Recognition Algorithm

Automatic facial expression recognition is an important research area for the computer vision field.It has huge application value in human-computer interaction,virtual reality,aug-mented reality,driver assistance systems,education and other aspects.With the rapid devel-opment of artificial intelligence,automatic facial expression recognition has been more widely studied in recent years.At present,there are many difficulties in deep learning-based facial ex-pression recognition approaches,such as training hard,limited migration platform and lack of database,etc.Therefore,how to ensure excellent recognition accuracy and conveniently train and transplant models has become a difficulty and focus of this topic.From the perspective of feature engineer,the recognition algorithm of facial expression in 2D video with traditional methods is proposed and verified the effectiveness and robustness of the algorithm with expres-sion videos recorded in wild environment.The Constrained Local Model is firstly summarized,and then the Deformation Model,Lo-cal Model and the fitting optimization are mainly introducedand.On this basis,the principle and characteristics of CLNF model and CE-CLM model are introduceed in detail.The detec-tion performance of several models on the frontal face and profile face in the unconstrained environment are explored through experiments on multiple face datasets.The experimental re-sults show that the detection accuracy of CE-CLM is very high and the detection of profile face image is more robust.Considering that the action units encoded in FACS can reflect the state of facial muscles,the intensity information of AUs is regarded as the basis for facial expression recognition here.To solve the problem of insufficient representation ability of single feature,geometric feature fuses appearance featurethis that directly composed of LBPH feature and HOG feature so as to achieve the purpose of complementing edge feature and local feature and improving the de-tection performance of AUs.Intensity estimation of AUs is regarded as a multi-classification problem and multi-kernel support vector machine is proposed to improve the classification ef-fect.Finally,the benefits brought by fusion features and MK-SVM in the detection system are explored through experiments and the results show that the model with them is superior to the reference model.Traditional facial expression recognition algorithms generally use spatial and frequency domain features to recognize expressions in 2D image sequences.From the perspective of in live detection,this paper studies how to use machine learning to extract features containing rich expression information from the intensity value of AUs.For this reason,a histogram statistical featurest is proposed to describe the expressions in the videos and the multi-class classifier com-bined with SVM and Adaboost algorithm for classification is applied.In BP4D-spontaneous database,k fold cross validation was used to compare the classification performance of geo-metric features,sequence features and statistical features in FER system,so as to validate the feasibility and effectiveness of statistical features.In order to explore the generalization ability of the model and the recognition accuracy in complex scenes,a small dynamic dataset is created and the results illustrate the system has high robustness.