Estimating 3D head poses accurately in low resolution video is a challenging vision task because it is difficult to find continuous one-to-one mapping from personindependent low resolution visual representation to head pose parameters. We propose to track head poses by modeling the shape-free facial textures acquired from the video with subspace learning techniques. In particular, we propose to model the facial appearance variations online by incremental weighted PCA subspace with forgetting mechanism, and we do the tracking in an annealed particle filtering framework. Experiments show that, the tracking accuracy of our approach outperforms past visual face tracking algorithms especially in low resolution videos.