Illumination Invariant Face Tracking and Recognition

Face tracking and recognition are not easy because the face is a non-rigid object. The main reasons for the failure to track and recognize faces are changes of pose and environmental illumination. In this paper, we propose a preprocessing method that compensates for illumination variations and a nonlinear manifold framework to deal with pose changes in the face. For video-based face recognition system, we embodied our system into two parts: the illumination preprocessing, and tracking and recognition modules. To handle illumination variations, we first preprocessed the input images to the single scale retinex images. And then we compensate the reflectance by applying histogram fitting on the defined range. To track and recognize the person, we approximate the face pose density by principal component analysis (PCA) using images sampled from training video sequences and then construct the Gaussian mixture model for each person. After that, we recognize and track the person using the maximized pdf from the approximated nonlinear manifold that is updated sequentially. Experimental results show that our proposed method is more efficient than the other ones for pose and illumination variations.