Unlike conventional video-based face recognition systems, in which the tracking and recognition are considered as two independent components, this paper presents a new integrated framework for simultaneously tracking and recognizing human faces. In this framework, tracking and recognition modules share the same appearance manifold. During training, because locally linear embedding (LLE) can detect the meaningful hidden structure of the nonlinear face manifold, LLE combined with K-means is employed to assign face images of every individual into clusters to construct view specific submanifolds. To improve the robustness of tracking and recognition, robust locality-preserving projection is developed to obtain linear subspaces that approximate the nonlinear submanifolds. Dynamics is also learned during this period. During testing, to reduce the great computational load, the integrated posterior probability is partitioned into two independent probabilities, which are obtained by a particle filter and by maximum posterior estimation by Bayesian inference, respectively. Extensive experimental results show that our proposed framework is effective for tracking and recognition under significant variations in pose, facial expression, and illumination and under scale variations and partial occlusion.