Tracking facial features using Gabor wavelet networks

The work presents a method for automatic facial feature tracking in video sequences. In this method, a discrete face template is represented as a linear combination of continuous 2D odd-Gabor wavelet functions. The weights and 2D parameters (position, scale and orientation) of each wavelet are determined optimally so that the maximum amount of image information is preserved for a given number of wavelets. We have used this representation to achieve effective facial feature tracking that is robust to homogeneous illumination changes and affine deformations of the face image. Moreover, the tracking approach considers the overall geometry of the face, being robust to facial feature deformations such as eye blinking and smiling. The number of wavelets in the representation may be chosen with respect to the available computational resources, even allowing real time processing.