A SURVEY ON MONOCULAR VISUAL ODOMETRY FOR AUTONOMOUS VEHICLE NAVIGATION

Abstract

This paper aims to study the current state of research in monocular visual odometry (VO), especially
towards autonomous vehicle navigation. It is necessary for a vehicle to keep track of its position in its
environment across the span of its journey. A technique that allows a robot to find its position and
orientation in space from an image stream constantly captured by a camera attached to it, is visual
odometry. Monocular sequences of image data obtained while driving contains lots of information, and
hence visual odometry is a viable solution for localization and has also proven to be robust. This paper
presents the existing study in VO, and compares the different techniques in detail, in terms of their
approaches. The state-of-the-art deep neural networks-based models are categorized and discussed.
Furthermore, potential areas of improvement and approaches that are yet to be implemented are
discussed.