Democritus University of Thrace (DUTH)
Methods to refine the self - localization of planetary rovers using orbital imaging
Abstract
dc:descriptionOne of the most common solutions for the exploration of remote and harsh environments such as volcanic or deep see areas is the employment of a robotic system. In the case of space exploration and specifically to areas which are further away from the orbit of Earth, the robotic exploration is the norm, considering that the human species has not yet been proven to be able to survive out of Earth's atmosphere -or a man made construction, such as the ISS- for extended time periods. Thus, the exploration of Mars has been performed -so far- exclusively by robots. Notwithstanding the good performance of contemporary space rovers, on autonomous or semi-autonomous operations, several issues must be assessed to allow more complex missions. The accurate localization and route estimation of space exploration rovers, without relying on a ``human in the loop", constitutes one of the most severe hindrances towards the success of large scale space exploration missions such as the construction of a habitat for humans. The contemporary rovers of the National Space Administration of USA (NASA) which currently explore Mars (namely the MER-A and MER-B of the mission Mars Exploration Rovers-MER and the Curiosity of the mission Mars Science Laboratory-MSL) navigate relatively to an arbitrary frame of reference (e.g. their landing point), ignoring the planet-specific coordinate frame.The motivation behind this PhD dissertation has been the prospect to utilize the new and high quality sensors orbiting Mars such as the High Resolution Imaging Science Experiment (HiRISE), in order to extract landmarks and eventually employ them to localize a rover autonomously on an absolute frame of reference. But, autonomy requires the researchers to follow an interdisciplinary approach. In the context of this research several algorithms have been developed -each on a different field- which where all connected in the end to produce a system capable of delivering the absolute position on Mars.This dissertation is articulated into six chapters, each of which tackles a separate subject. Let us begin with chapter one which includes all the basic understanding and motivation required to follow this work as well as a literature survey that presents all the previous works on the subject. Concerning the literature survey, a clear taxonomy of the approaches has been performed based on the type of sensory each of the methods follows as well as based on the approach employed. Following the conclusions of chapter one, chapter two elaborates on the crucial subject of the employment of proper and realistic datasets. The first dataset utilized, is the one that was produced by ESA in the context of an activity where a robot was employed to navigate in the Desert of Attacama. Indifferently to the alikeness of the Atacama desert to a typical Martian area, we decided to replicate -in terms of boulder distribution- an actual area on Mars, namely the Gale Crater. With a view to provide proper data we constructed a robot (Heavy Duty Planetary Rover) the vision system of which is a replica of the one on the ExoMars rover. The dataset produced, is not only valuable for us, but will also benefit other researchers as we provide it freely through ESA's web servers.The third chapter includes all the methods required to provide accurate relative localization. Even though relative localization -through visual odometry- produces an error which is unbounded, the error's growth rate is mostly important. We developed two approaches, one that is on par with the current rover practices and a second one, which is more advanced, yet still comparable to the first one in terms of computation effort demands. The first algorithm utilizes the 3D reconstruction of a stereo pair -onboard the rover- and along with the 2D matching procedure, provides 3D matches which are refined through a RANSAC procedure. Then a PnP problem is solved through the EPnP method, the output of which is loosely coupled with an IMU through an EKF. The second approach we developed employs a window based Bunlde Adjustment, with the window been variably adjusted according to the indications of an accurate IMU. The results produced by both of these two algorithms are adequate and on par with the state of the art. There are several parts of these two algorithms that to the best of our knowledge have never been implemented before, such as the IMU control of the size of the Bundle Adjustment's window.The fourth chapter describes both the design and the evaluation of a custom pipeline that we employ to autonomously detect and classify Regions of Interest (ROIs) on images stemming from rovers traversing a planet and from orbiters that perform observation. The detection of possible ROIs is performed on simple, yet effective, computer vision metrics. The Machine Learning (ML) approach, is the one of the binary classification for each type of landmark. The training of the system is performed offline while the classification is autonomously executed on-line. We decided to tune and benchmark a plethora of classifiers -namely ``K Nearest Neighbors" (k-NN), ``Naive Bayes" (NB), ``Logistic Regression" (LR), ``Classification Tree" (CTree), ``Support Vector Machine" (SVM) and ``Gaussian Process" (GP). Finally, in the fifth chapter the methodologies that allow us to build and match sparse networks from rovers and orbiters are presented. Two networks dominate our global localization approach, namely the Local Network (LN) and the Global Network (GN). The LN is created using the landmarks which the rover has detected and classified, while the GN is built upon a georeferenced -or areoreferenced- image. The system offers two modalities for the global localization. The first is the semiautonomous DARCES approach, which is able to provide sorted solutions for the rovers localization. The second approach is a completely autonomous one that employs graph optimization. The graph is populated as a normal pose graph, except from the fact that the local and global part of the graph are connected only through ``loop closures" and not relative localization. We employ an approach able to handle false positive, so we can overcome the problem of DARCES providing more than one solution.This doctoral dissertation is finalized with its last chapter. It first sets the basis for discussion concerning the extraterrestrial global localization of mobile robots during the last 30 years. Then it presents the approach followed in this work and our successful research achievements. Finally our idea, of the short and long term additions can be implemented in our system and how this work can be useful to other institutes and subjects, is presented.The contribution of this dissertation to the state of the art is clear and proven by our publication list. Prior to our research efforts the literature clearly lacked any method to provide accurate global localization employing the semantic characteristics of landmarks seen in both rover and orbital imagery. In particular, the system proposed within the framework of the thesis at hand includes: ``vision based motion estimation (relative localization), ``computer vision detection of landmarks", ``machine learning for landmark classification", ``graph optimization for pose and route refinement". All these components per se, as well as the compounded system, lay at the edge of the state of the art, since they match and/or exceed any previous work on the respective subjects. Concerning the relative localization we have implemented a EPnP -non iterative- solution with a variable window bundle adjustment, with the window being adjusted on-line based on an accurate IMU. This is the first time such a pipeline scheme has ever been presented in the literature. Concerning the classification of Regions of Interest -for the scenario of a planetary exploration mission- and after a thorough evaluation of the related work we were unable to find any research that reported the methodology followed in order to select the most efficient approach for the recognition of different types of ROIs, on rover and orbital images. Therefore, we conducted for the first time such an evaluation with remarkable results.The usage of orbital and rover ROIs in a graph to solve the problem of absolute localization of a robot in a ``lost-in-space" scenario has never been reported in the past. The technology developed in the context of this research will advance both future and currently planed missions. Concerning researchers currently localizing rovers on Mars manually, instead of searching an immense area they will, from now on, only need to search a handful of positions. Also, the autonomous capabilities of our system will push the envelope for future space robotic missions, such as Mars Sample Return and human planet colonization.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2016
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Boukas, Evangelos
- Μπούκας, Ευάγγελος
Subjects
dc:subject × 16- Ρομποτική
- Ρομποτική όραση
- Μηχανική μάθηση
- Διαστημική ρομποτική
- Μηχανική εκμάθηση με χρήση εικόνων
- Μηχανική όραση
- Μηχανοτρονική
- Robotics
- Machine learning
- Space robotics
- Vision based machine learning
- Computer vision and image processing
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/39135