Active Appearance Model (AAM)
     In face recognition application, accurate face alignment has determinative effect. Active Appearance Model (AAM) is one of the most studied methods for accurate locating objects. When applying Active Appearance Model, firstly we collect enough face images with various shapes as training set. Then we use a set of points to annotate face shape, so face shape can be represented by the coordinates of these landmarks.
                               Construction of mean shape
     After a series of transformation such as Principal Component Analysis, the mean shape of all the faces can be obtained to construct shape model for face alignment.
    Given a new face image, we estimate the model’s initial position, compute the suggested movements, then we can get a good face alignment result.
Example of face alignment

    We mainly use shape parameters and appearance parameters obtained by alignment for real time video tracking and pose estimate. The speed of this algorithm is 123 frames/second. So it is applicable in real time face tracking. The pose estimate we implement contains: horizontal and vertical movements; forward or backward from camera; rotation angle; eyes and mouth state.

State estimate using AAM

 

Example of real time tracking and pose estimate
                                            Result analysis of one frame
The application using AAM in the car

Using AAM, the driver’s state can be detected and analyzed.