A framework for efficient real-time face servoing

The aim of this paper is to present solution for face servoing problem with an active camera. The algorithm is easy to implement, requires neither calibration nor expensive equipment. It performs with satisfactory degree of efficiency and promises wide range of applicability. The algorithm consists of two main parts: face detection and control part. Skin color distribution is modeled in CbCr space (derived from YCbCr) with multivariate Gaussian function. Squared Mahalanobis distance is used as a measure of similarity between image pixels and statistical skin prototype. After that step, analysis of geometrical properties of connected components is performed within skin detection mask, for the purpose of finding the best candidate(s) for the face. When the face is found, centroid vector is fed to Kalman filter. Error signal is formed as difference between filtered centroid vector and the center of image. Two decoupled PD controllers generate control signals for pan and tilt servo motors that drive the camera. The performance results of the algorithm, running on a real-time system, are presented within this paper.