Odometria visual monocular em robôs para a agricultura com camara(s) com lentes "olho de peixe"

Abstract

One of the main challenges in robotics is to develop accurate localization methods that achieve acceptable runtime performances. One of the most common approaches is to use Global Navigation Satellite System such as GPS to localize robots. However, satellite signals are not full-time available in some kind of environments. The purpose of this dissertation is to develop a localization system for a ground robot. This robot is inserted in a project called RoMoVi and is intended to perform tasks like crop monitoring and harvesting in steep slope vineyards. This vineyards are localized in the Douro region which are characterized by the presence of high hills. Thus, the context of RoMoVi is not prosperous for the use of GPS-based localization systems. Therefore, the main goal of this work is to create a reliable localization system based on vision techniques and low cost sensors. To do so, a Visual Odometry system will be used. The concept of Visual Odometry is equivalent to wheel odometry but it has the advantage of not suffering from wheel slip which is present in these kind of environments due to the harsh terrain conditions. Here, motion is tracked computing the homogeneous transformation between camera frames, incrementally. However, this approach also presents some open issues. Most of the state of art methods, specially those who present a monocular camera system, don't perform good motion estimations in pure rotations. In some of them, motion even degenerates in these situations. Also, computing the motion scale is a difficult task that is widely investigated in this field. This work is intended to solve these issues. To do so, fisheye lens cameras will be used in order to achieve wide vision field of views.