This paper presents a comparison of natural feature descriptors for rigid object tracking for augmented reality (AR) applications. AR relies on object tracking in order to identify a physical object and to superimpose virtual object on an object. Natural feature tracking (NFT) is one approach for computer vision-based object tracking. NFT utilizes interest points of a physcial object, represents them as descriptors, and matches the descriptors against reference descriptors in order to identify a phsical object to track. In this research, we investigate four different natural feature descriptors (SIFT, SURF, FREAK, ORB) and their capability to track rigid objects. Rigid objects need robust descriptors since they need to describe the objects in a 3D space. AR applications are also real-time application, thus, fast feature matching is mandatory. FREAK and ORB are binary descriptors, which promise a higher performance in comparison to SIFT and SURF. We deployed a test in which we match feature descriptors to artificial rigid objects. The results indicate that the SIFT descriptor is the most promising solution in our addressed domain, AR-based assembly training.
Online learning and fusion of orientation appearance models for robust rigid object tracking
