Fiducial facial points tracking using particle filter and geometric features

Tracking of facial points in image sequences has become an important and challenging task in the field of computer vision. In this paper, particle filters are used for simultaneously tracking of 18 points around intransient facial features (mouth, eye, eyebrow and nose). There are two stages for tracking of points. In the first stage, each point is tracked independently by a SIR particle filter. As there is interdependency among points of each intransient facial feature, in the second stage, for increasing accuracy of the facial points tracking, geometric features between facial points are applied on the output of the first stage. Nostrils usually don't have non-rigid changes when a facial expression is occurred on the face, therefore these two points result in more accurate tracking in image sequences compared to other facial points and hence are used as reference points in each frame. The proposed method, unlike some previous works, doesn't require any training and therefore is fast and not dependent on the used data set. The proposed algorithm is applied on Cohn-Kanade database and experimental results are presented to show the effectiveness of the algorithm.