Date of Award

3-10-2009

Document Type

Thesis

Degree Name

Master of Science in Electrical Engineering

Department

Department of Electrical and Computer Engineering

First Advisor

Michael J. Veth, PhD

Abstract

The Air Force Institute of Technology (AFIT) is investigating techniques to improve aircraft navigation using low-cost imaging and inertial sensors. Stationary features tracked within the image are used to improve the inertial navigation estimate. These features are tracked using a correspondence search between frames. Previous research investigated aiding these correspondence searches using inertial measurements (i.e., stochastic projection). While this research demonstrated the benefits of further sensor integration, it still relied on robust feature descriptors (e.g., SIFT or SURF) to obtain a reliable correspondence match in the presence of rotation and scale changes. Unfortunately, these robust feature extraction algorithms are computationally intensive and require significant resources for real-time operation. Simpler feature extraction algorithms are much more efficient, but their feature descriptors are not invariant to scale, rotation, or affine warping which limits matching performance during arbitrary motion. This research uses inertial measurements to predict not only the location of the feature in the next image but also the feature descriptor, resulting in robust correspondence matching with low computational overhead. This novel technique, called deeply-integrated feature tracking, is exercised using real imagery. The term deep integration is derived from the fact inertial information is used to aid the image processing. The navigation experiments presented demonstrate the performance of the new algorithm in relation to the previous work. Further experiments also investigate a monocular camera setup necessary for actual flight testing. Results show that the new algorithm is 12 times faster than its predecessor while still producing an accurate trajectory. Thirty-percent more features were initialized using the new tracker over the previous algorithm. However, low-level aiding techniques successfully reduced the number of features initialized indicating a more robust tracking solution through deep integration.

AFIT Designator

AFIT-GE-ENG-09-17

DTIC Accession Number

ADA499487

Download

Share

COinS