Simple feature identification and matching for stereoscopic imaging

Abstract

Many camera calibration techniques used in stereoscopic imaging require corresponding points in image pairs to be identified. The best matches tend to come from very distinctive regions, hence it is preferable to use only these points and reject points from more homogeneous regions. Often points are selected and matched by a combination of user input and automated matching. However, for most techniques, the more matching points identified, the more accurate and robust the calibration. Therefore, it is desirable to alleviate the need for user input and to automate the point selection process. In the feature matching technique described here, a combination of object extraction, feature identification and feature matching is used. This technique is simple, efficient and robust. The algorithm is to be used as part of a larger project in 3D human motion capture which is currently under development at University College Dublin and the National Rehabilitation Hospital in Dun Laoghaire.