Real Time Head Tracking From Uncalibrated Monocular Views

摘要

In this paper, a model-based head tracking from monocular and non-calibrated video sequences is presented. The proposed method relies on the matching of a 3D generic head model and 2D image features extracted from the input sequence. Head tracking is based on the minimization of an error function which describes the discrepancies between model and image features. Motion and texture information are considered in order to make tracking stable. Minimization is obtained applying a gradient based technique. The sequence of reconstructed head poses allows simple gesture recognition. After pose reconstruction, the input image is warped into the texture map of the model. The stabilized view of the face obtained, can be used to improve facial expression analysis and reconstruction. The overall performance of the non-optimized head tracking algorithm is about 30 frames/sec on a Pentium III 500. Data about the reconstruction accuracy achievable with our technique are also presented.