A Multi-Camera Pose Tracker for Assisting the Visually Impaired

6DOF Pose tracking is useful in many contexts, e.g., in augmented reality (AR) applications. In particular, we seek to assist visually impaired persons by providing them with an auditory interface to their environment through soni.cation. For this purpose, accurate head tracking in mixed indoor/outdoor settings is the key enabling technology. Most of the work to date has concentrated on single-camera systems with a relatively small .eld of view, but this presents a fundamental limit on the accuracy of such systems. We present a multi-camera pose tracker that handles an arbitrary con.guration of cameras rigidly .xed to the object of interest. By using multiple cameras, we increase both the robustness and the accuracy by which a 6-DOF pose is tracked. However, in a multi-camera rig setting, earlier methods for determining the unknown pose from three world-tocamera correspondences are no longer applicable, as they all assume a common center of projection. In this paper, we present a RANSAC-based method that copes with this limitation and handles multi-camera rigs. In addition, we present quantitative results to serve as a design guide for full system deployments based on multicamera rigs. Our formulation is completely general, in that it handles an arbitrary, heterogeneous collection of cameras in any arbitrary con.guration.