3D real-time facial feature points tracking with improved particle filter

Real-time face alignment is very crucial in many applications, such as performance-driven facial animation and face recognition. However, traditional face alignment techniques are only good at dealing with static pictures. Little attention has been paid to face alignment in video. In this paper, we present a novel framework named particle filter based AAMs (PF-AAMs), for retrieving the parameters of the most-well known face method, Active Appearance Models (AAMs), to deal with real-time facial landmark tracking problem. The proposed algorithm has two contributions. First, the statistic property of particle filter algorithm can greatly handle many common problems in gradient descendant algorithm. Second, in order to solve the difficulty of particle filter algorithm in high-dimensional optimization problem, an improved particle filter method is proposed in real-time scene. Experiments on YouTube faces database indicate that the proposed method obtains better performance than AAMs with fewer computational cost.