Facial Feature Points Detection And The Application In Facial Animation

Facial feature points detection is a basic research in the field of machine learning,that widely used in face data analysis,face recognition and face reconstruction.The major difficulty of existing methods about facial feature points detection is that dramatic changes in lighting,facial expressions,gestures and other factors caused loss of characters,eventually leading the failure of facial feature points detection.Typically,popular facial feature points detection algorithm aims at frontal human faces for training and testing,fails to multi-angles facial feature points detection.However,many real-life applications belong to the research of multi-angles facial feature points detection,such as criminal identification,detection of sensitive people,facial recognition and facial animation in unlimited scenes.Multi-angles head posture would lead to non-linear changes in facial appearance,affecting facial feature points location greatly.Due to the impact of these external factors,the precision location of facial feature points is a challenging research tasks.This paper analysis and research the existing algorithm,proposes a improved conditions regression forest based on the head pose,The algorithm has change the performance of nonlinear problem into a linear one.The main work of this paper consists of two parts.The first part is for head pose estimation,we use locality preserving projection(LPP)and generalized regression neural network(GRNN)for obtaining the global information of pose and label it,the main idea is used manifold learning for image data dimensional reduction,then mapped the data into linearly separable space via nonlinear regression,then using the results of nonlinear regression for the head pose estimation.Experiments show that our method could estimate the head pose of images better,with fast speed and high robustness.The second part is for facial feature points detection,The paper proposes a conditional regression forest to estimate facial feature points in two-dimensional images in real-time based on the estimated head pose.Some changes of the face are depend on global facial feature,such as head pose,so the paper categorized human faces base on head pose and then learning separately.Conditional regression Forests are get several conditions from the parameter space instead of conventional random regression forests with a single probability,so unnecessary to deal with all thechanges in the appearance of the face and shape.Through a variety of different head pose information corresponding to each training condition tree,we also choose corresponding condition trees that with the same head pose in training for prediction.Experimental show that the method effectively reduces the non-linear estimation accuracy error caused by the change in image space and other features of the head deflection,facial feature points location been increased,speed up to 30 fps and meet the real-time requirements.Finally is facial animation synthesis,analysis the information based on the estimated facial feature points,set up the relationship between images and three dimensional model.And then training the radial basis function regression network by this relations,building the facial animation synthesis system.The paper using Microsoft Visual 2010 development tools and then drive Ogre model for animation synthesis.The system is not only the embodiment instantiation of our methods,but also provide a technology platform for future research.