Tracking Head Pose In Unconstrained Environment

摘要

In this paper, we present an approach to track and estimate the head pose of a people. The motivation of our work is to build a vision interface for a mobile agent system. In such applications, it is not feasible to assume the environment as a constrained one, which most previous approaches assumed implicitly. We use color as basic cue to detect face from image. Color model is initialized interactively when systems starts and is adapted online during the tracking process. View-verified scheme is adopted to cope with the problems arise from the complex background. For each extracted face image, a histogram-normalized grayscale image and corresponding edge image are used to train a self-organizing map (SOM) for estimating the head pose. The experimental results using data taken from different environments show that this approach is both stable and good in accuracy