Deep-Learning-Based Visual Odometry Models for Mobile Robotics

Abstract

Odometry is a common problem in navigation systems where there is a need to estimate the position of the vehicle or carrier in the environment. To perform autonomous tasks, robotic or intelligent devices need to be aware of their position in the environment. There are many strategies to solve an odometry problem. This work explores a visual odometry solution with a deep neural network to infer the robotic vehicle's position in a known and mapped environment. The first robot, equipped with a LIDAR, IMU, and camera, maps the environment through a SLAM technique to perform this task. The data gathered by this first robot is used as ground truth to train the neural network, and later, other robots with only one camera can locate themselves in the environment. We also propose a validation and evaluation of the neural network.