Real-time Facial Features Tracking by Discrete Gabor Jets and Mean Shift

This work presents a novel, real-time facial features tracking algorithm. In the extraction step angular and radial frequencies in circular regions around fiducial points are evaluated using approximation of the Gabor filter - discrete Gabor jets. Within this approach only few simple operations followed by the fast Fourier transform are performed instead of highly time consuming Gabor filter calculations. Classification is performed using a modified linear discriminant analysis adapted to the facial/non facial feature class problem. Mean shift algorithm is used to find local maxima for the set of the found features. Only edge points, evaluated using the Sobel operator and non-maximum gradient magnitude suppression, are considered as fiducial points. Accuracy of the algorithm is analysed with respect to the cut-off threshold of the gradient magnitude in the edge detector and distance threshold to the LDA model in classification procedure.