Shape recognition and tracking for augmented reality

Abstract

The field of 3D vision still provides many challenges in research. This thesis

discusses 3D computer vision in real-time tracking and recognition for augmented

reality to seamlessly merge 3D virtual objects into real image scene captured by a

camera. The recognition is based on the shapes of detected objects and a set of related 2D templates. The correspondence problem is solved in a top-down recognition

framework using model-based detection and tracking from 2D views. Tracking

flexibility is increased for wide base line matching by using the contour shape of

detected objects. The cross correlation of r-signature between the object shape and its

corresponding template is performed to improve the contour detection. Subsequently, it is verified by the planar template reprojection using the homography transformation to get the 3D pose parameter. A case study is applied to alphabetic letter recognition. The angle view effect on the cross correlation value is evaluated and stabilized by using the aspect ratio normalization.