A Framework for the Design and Implementation of a Dynamic Face Tracking System

This paper presents a 3-tier dynamic face tracking system (DFTS) framework based on Gabor wavelets. The proposed system is composed of three major modules: face detection unit, Gabor decomposition & similarity computation unit and face-lock refinement unit. First, we present a neural network based face detection that is used for face localization in cluttered environments. Subsequently, Gabor decomposition using a Gabor wavelet filter bank with 40 filters (8 orientations and 5 scales) of the initial image with the detected face, is carried out for facial feature point (FFP) extraction and initialization. In the next stage, similarity algorithms along with gabor wavelet decomposition are used to track the movements of FFPs in the faces. Finally, a face-lock mechanism using a subset of MPEG-4 facial definition parameters (FDP) is utilized to fine-tune the face tracking system under partial occlusion conditions. Most of the current face tracking systems need the FFPs to be labeled manually. They assume that the targeted face has already been detected and localized. In our proposed approach, face detection followed by automatic initialization of FFPs using Gabor wavelets is done to facilitate face tracking in a seamless manner. Rapid prototyping of FPGAs to speed up the compute- intensive tasks was performed.