Localização baseada em odometria visual para robôs móveis terrestres

Abstract

This work proposes and evaluates a localization approach for ground mobile robots based on Visual Odometry, with a focus on low-cost systems and widely available sensors. The methodology estimates the robot’s pose from images captured by a monocular camera, using geometric reconstruction techniques, triangulation, and optimization via Bundle Adjustment. To address the limitation of monocular Visual Odometry regarding metric scale recovery, a Wheel Odometry system is also implemented, based on encoder measurements. The Visual Odometry and Wheel Odometry estimates are then compared and evaluated both quantitatively and qualitatively, using standard metrics such as Absolute Pose Error and Relative Pose Error, based on reference trajectories from public datasets and a real robot. The results show that Visual Odometry can outperform Wheel Odometry in accuracy in several scenarios, especially in environments with slippage or irregular terrain, highlighting its potential as a key component in hybrid localization systems.