An interactive vision-based tool for model-based scene calibration of augmented reality environments

Abstract

Non-trivial Augmented Reality environments consist of a recon gurable set of objects whose spatial relations need to be calibrated upon setup. Since manual measurement is tedious, timeconsuming, and error-prone if not done carefully, fully or semi-automatic procedures are needed. We have developed a semi-automatic vision-based scene calibration tool which allows a user to perform this task e ciently and accurately. For each scene object to be calibrated, the user interactively speci es its approximate pose by aligning a 3D model of the object with its projection in one or several camera image(s) of the scene. Thereupon, the exact pose is automatically computed by a combination of computer graphics and computer vision techniques. The combined knowledge of the object model and of the initial pose estimate allows the system to signi cantly improve the feature matching and pose estimation processes. Because the latter is based on a robust estimation method, the system performs well even in cluttered environments where objects are partly occluded.