Research On The Robust Methods Of Facial Feature Points Detection And Tracking With Various Expression

With the development of computer technology, artificial intelligence has been widely used to the our lives. As an important part of face intelligence analysis, facial feature detection and tracking can get accurate information of local parts of human face, which plays an important roles in the further research. In this paper, a method of feature points location in static images and a method of tracking in video sequence are proposed under facial expressions. The experimental results show that the proposed detection and tracking method shows good robustness under the corresponding conditions.In order to solve the problem of facial feature points localization in images, a new feature location method based on multiple feature and multiple kernels learning is proposed. First,using a series of face database that has been marked landmarks to create a shape statistical model. After the initialization is completed, the built-in local detector which are trained using multiple feature and multiple kernels learning methods is used to detect and output the response graph at the local area around each initialization point. Finally, we combine the quadratic function on the local response graph with the global model constrained iterative optimization function to complete the feature point detection. Experiments show that the method has achieved better localization effect in static images under facial expressions, and the accuracy is improved obviously compared with the feature point detection method of support vector machine based on single feature single kernel.In order to solve the problem of facial feature points tracking in video, an improved spatio-temporal context tracking method is proposed, and the feature point tracking is completed by the local limit model feature point detection method. First, using the feature point method to give the initial position. Second, the learning space context and the time context produce a confidence map. Again, the higher confidence level is calculated in the confidence graph and the similarity coefficient of the current prediction region and the previous frame target region is calculated. Finally, it is judged whether the similarity coefficient satisfies the set threshold. If it is not satisfied, the iteration calculation is repeated with the current predicted position until the threshold of the similarity judgment coefficient is required to stop the iteration and output the tracking result. By introducing iterative optimization process, the accuracy of target tracking is improved. The experimental results show that the proposed method can track the feature points of the images with different expressions in the video sequence.