Urban navigation based on visual odometry for people with low vision

Abstract

This document includes and explains the final project of the master’s degree in Automatic Control and Robotics. The aim of this is the creation of an algorithm that improves the location of individuals with low vision. The methodology used consists of the realization of several successive phases that, in a final instance, allow to reach the objective set. First, research is carried out on the different navigation and location solutions that exist today. Next, three localization correction methods are proposed, based on visual odometry, cartographic information and image segmentation. Using these data, the three models proposed are based on limiting the location to the transitable areas of the map. Correct the location on the sidewalk using the lateral distances at the ends of it. And calculate the rotation and translation necessary to align the margins of the cartographic sidewalk with the segmented mask from the real images. Due to the concretion of the problem raised, a database with sequences transiting sidewalks has been created to evaluate the developed methods. To validate the different methods, recorded sequences have been used, and different metrics have been evaluated for each of the sequences. Finally, the obtained result presents different methods that allow the improvement of the location, although it opens the doors to continue developing solutions that improve not only the correction itself, but the data used for it.