Wearable visual sensors provide views of the environment which are rich in information about the wearer s location, interactions and intentions. In the wearable domain, hand gesture recognition is the natural replacement for keyboard input. We describe a framework combining a coarse-to-fine method for shape detection and a 3D tracking method that can identify pointing gestures and estimate their direction. The low computational complexity of both methods allows a real-time implementation that is applied to estimate the user's focus of attention and to control fast redirections of gaze of a wearable active camera. Experiments have demonstrated a level of robustness of this system in long and noisy image sequences.