Research On Algorithms For Facial Feature Point Tracking

With the rapid development of information technology and artificial intelligence, computer has the ability to simulate human thinking and wisdom. Computer vision has become one of hot research topics. In recent years, face detection, face recognition, feature extraction and facial feature point tracking methods are becoming much richer. Facial feature point tracking is a basis for the subsequent face recognition and facial expression analysis. It has broad application prospects. Researchers have proposed many facial feature point tracking algorithms. However, in real applications, because of the variation of the face posture, occlusion, facial expressions and other internal and external factors, it makes the facial feature point tracking become more and more complex. Thus, the issue of facial feature point tracking with efficiency and robustness deserves more research than it has received.The main research and innovation of this paper is reflected in the following aspects:First, we reviewed the existing face detection algorithms at home and abroad, then we use the Viola-Jones face detector and we do detailed theoretical description and analysis of it. Then we do experiments to show the effectiveness of the Viola-Jones face detection algorithm.Secondly, we show the Active Appearance Model(AAM) and the Constrained Local Model(CLM), we do a detailed theoretical introduction and comparative analysis for them. Then we do experiments to compared the AAM-based Lucas-Kanade reverse algorithm and CLMbased Regularized Landmark Mean-Shift and analyze the advantage of using CLM model in real-time facial feature point tracking.Final, we improved the CLM in many aspects. The application of CLM would cause the tracking failure of facial feature point when face pose greatly deflected and initial position deviated drastically from the target face. To solve this problem, we use the probability model of Isotropic Gaussian kernel density estimation. We also calculate the Mean-Shift vectors to ensure the accuracy of track result. M-Estimation is used to overcome interference of the abnormal values when partial occlusion occurs.Meanwhile, limitation of the feature points search area and approximate calculation method ensures the real-time operation of the system.The experimental results show that the proposed algorithm has both efficiency and robustness.