Affine Feature Matching via Stochastic Prediction

Kenneth A. Fisher, Jared Kresge and John F. Raquet

Abstract:	Tracking features from visual sensors for navigation purposes has emerged as a promising augmentation to convention sensors such as inertial measurement units (IMUs) and the Global Positioning System (GPS). In a tightly-coupled Extended Kalman Filter, errors may be reduced by approximately two orders of magnitude [11]. While the basic mathematics and algorithms are thoroughly documented, image-aided navigation is still in its early stages. This research improves image-aided navigation's feature tracking and landmark database by improving feature matching and landmark characterization across a wide range of viewpoints. In particular, current feature descriptors are typically based upon a Scale Invariant Feature Transform (SIFT) [5] and can only be matched reliably when viewed within ±30 degrees of the original viewpoint [1]. In this paper, it is shown experimentally that stochastic affine prediction expands the viewpoint validity to ±60 degrees. Furthermore, this description improves landmark databases by including a viewpoint dependency. Using real-world data, affine feature matching via stochastic prediction reduces navigation errors by 24% in position and 35% in attitude compared to the standard two-camera image-aided navigation setup.
Published in:	Proceedings of the 25th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS 2012) September 17 - 21, 2012 Nashville Convention Center, Nashville, Tennessee Nashville, TN
Pages:	660 - 669
Cite this article:	Fisher, Kenneth A., Kresge, Jared, Raquet, John F., "Affine Feature Matching via Stochastic Prediction," Proceedings of the 25th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS 2012), Nashville, TN, September 2012, pp. 660-669.
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In