On Monocular Visual Odometry for Indoor Ground Vehicles

Abstract

In this work, a Visual Odometry system estimating camera position and orientation based solely on image meausurements is proposed. The system is built on the fundamentals of Structure from Motion theory, requiring only a single calibrated camera to estimate positional information. Experiments were conducted on publicly available datasets with associated Ground Truth, in which a consumer grade web camera was employed as the system's sensor on realistic indoor scenarios. The proposed system resulted in absolute localization error rates ranging from 2.10% to 3.39% of the travelled distance at processing rates of up to 5Hz, performing better than wheel odometry in certain situations and being equip arable to similar approaches present in the robotics literature.