Towards More Robust Vision-based Map Matching Through Machine Learning & Improved Feature Matching

Abstract

Currently GPS is the primary source for all navigation solutions in the US military as it is relatively cheap and easy for a user to implement and readily available worldwide. However, the heavy reliance and investment in GPS creates several vulnerabilities in our military. GPS is easily jammed and manipulated by intentional and accidental means, which creates a need for other alternative navigation methods. One such alternative is computer-aided visual navigation. A main hurdle up to this point has been to establish meaningful and robust features for consistent and accurate feature matching between aerial and satellite imagery. Typical feature matching on aerial imagery results in a majority of features being placed on trees and other non-seasonally invariant features. The primary objective of this research is to test the effectiveness of using semantic segmentation as a way to create and force robust features onto desired areas of an image for the purpose of visual navigation. This involves testing several segmentation algorithms to achieve state-of-the-art segmentation results and evaluating the effectiveness of feature matching on segmented imagery. The focus of this research will be on the development of a near state-of-the-art semantic segmentation model for aerial imagery that can extract desired buildings from an image. The research will then focus on various feature selection and feature matching algorithms in order to compare the segmented aerial key features with a known database of features from satellite imagery. Current results show that feature selection algorithms such as SIFT fail to overcome the nuances between multi-source aerial imagery. Improving the feature selection algorithm will ideally allow for an increased quantity and quality of matches, ultimately resulting in a camera pose estimation sufficient to be a reliable alternative to GPS.