Abstract
Most of the facial animation applications, such as automatic face detection, recognition, and tracking are sensitive to the different lighting conditions and worse background environment, especially accurate expression analysis in virtual conferencing system. The main work in our system is to detect the face under these complex situations and reduce the influence of different lighting. In addition, the pose estimation is the bottleneck of the framework, the speed of which is to be improved.First, we adopt a skin color model, which includes the training data under different lighting conditions, and the pixels which belong to skin color are detected correctly in most situations. Instead of overall scanning the original image, we scan the down-sampled skin map and apply the skin mask to obtain the possible facial blocks fast. Further, a modified mean shift algorithm is used to refine the accuracy of overall scanning and skin mask and also improve the speed by increasing the step size of scanning. The lighting distribution is estimated by second order function, and the lighting effect is compensated. After lighting compensation, we introduce the support vector machine classification to check if a block includes face or not, and it is easier to classify the blocks of face or non-face in the candidates of the subset of skin color. Finally, we modify the good features selection algorithm to the texture map to pick out the features which contain the edges of two different directions, and then perform the analysis-by-synthesis pose estimation with the texture map of good features to speed up without losing accuracy.