Computer vision system for head movement detection and tracking

The increased popularity of applications requiring head tracking, such as teleconferencing and virtual reality, have fuelled research efforts to provide computer vision solutions to the problem of real-time head movement tracking. The attractiveness of this type of solution rests on the fact that head tracking can be performed without the use of expensive and cumbersome physical devices. We propose a computer vision approach that detects the head movements of a user seated at a computer workstation. We model head translation and head rotation using distinct sets of templates synthesized from an initially captured image of the head and representing this head in various positions (and sizes) and orientations. Using correlation-based template matching, we achieve detection by correlating these sets of templates against each image of the head captured by a camera positioned on the top of the monitor. The best-correlating template from the set modelling head translation and the best from the set modelling head rotation represent good approximations of the three-dimensional position and orientation of the head in the scene, respectively. We improve on these approximations by defining two functions that interpolate the correlation scores of each set and by obtaining the minimum of each of these functions. We use these two minima, which represent head position and orientation, to synthesize a new template based on the initially captured image of the head. This new synthesized template represents the image of the head that most closely approximates the head position and orientation in the scene. Head movement tracking is performed by comparing the closest approximation of the head found in two consecutive images of the scene. We have implemented our head movement tracking approach and found our system to track head position, on average, to within one pixel of the measured head position in both the x- and y-axis directions, and to detect head size (width and height), on average, to within one and two pixels of the measured head width and height, respectively. Our head movement tracking system tracks head rotations, on average, to within 1.4° of the measured angles.. Our tracker is capable of processing up to eight captured images per second.Science, Faculty ofComputer Science, Department ofGraduat