In this paper, an active vision system used for real time facial features tracking is presented, where a human head region is searched using a stereo camera. This region is divided such that a person's nose will be searched in a smaller spatial domain. The 2D coordinates of the nose are determined based on a Haar classifier and colour segmentation. Its 3D coordinates, computed via the constraints of the epipolar geometry, represent a reference signal used within a control system for controlling the stereo camera's pan and tilt. A method for determining the system's parameters and for designing a proper controller is further proposed. The controller is design such that it integrates the maximum time delay that can be inserted into the control system by the machine vision component. Finally, the overall architecture is tested in a real facial feature tracking application.