National Technical University of Athens (NTUA)
Ταξινόμηση κατάστασης θέσης σε γραμμικά γεωμετρικά πρότυπα με μηχανές επιβλεπόμενης μάθησης σε ζητήματα πλοήγησης και καθοδήγησης
Abstract
dc:descriptionConcretely, guidance entails not navigation alone, but mainly the utilisation of the navigation solution for the maintenance of a moving vehicle course from one location to another. Technically, it depends heavily on the trajectory determination in 3D space which in effect, it concerns with two aspects; firstly, location information expressed at a suitable coordinate system and secondly, geometric information that shapes the path of the moving vehicle in a formation of a sequence of mathematical curves. Location information results directly from the navigation solution in a form of individual, consecutively distributed point fixes (position estimates), whereas geometric information can be extracted using various modelling approaches. Notwithstanding position estimation can be performed in real time, trajectory geometry is essentially determined post-hoc, as a real time procedure would infer trajectory geometry identification requiring prior knowledge of the curvature along the path nonetheless. In this frame, the following research venture discusses questions upon guidance, an interdisciplinary subject also known as the art of navigation.More specifically, it proposes a methodology to perform geometry identification of a vehicle trajectory in real time. It is a work of synthesis of supervised machine learning and optimal state estimation for the real time classification of point fixes into appropriate geometric classes. On the basis of a categorisation axiom derived from attitude estimation, a Least Squares Support Vector Machine (LS-SVM) for pattern recognition classifies sequential position estimates into linear elements (such as straight lines, simple circular curves or lines of a curvature variation). Thus, the geometrical profile of the trajectory path is being identified while an Extended Kalman filter aboard a tightly coupled GNSS/INS system estimates vehicle kinematics. The elemental outcome from the application of the proposed methodology is the statistical estimation with which consecutive point fixes are constrained to belong to the same class of a linear pattern beforehand. Thence, the parametric equation of such a geometrically recognised line is simultaneously updated, as its fitting parameters are being recalculated after the classification of a successive point fix at the same class.The training scheme is designed on learning features created from the position-to-position time-series of attitude delta values. Given that the value of the trajectory curvature is a priori unknown, the conception of this simple trick offers an implicit quantification of the curvature rate of change, delivering an intuitive solution to tackle this point at issue. Inherently, a great single asset of the proposed methodology is centred upon the learning features the classifier is trained on; thus classification prediction is less influenced by the quality of the navigation solution. Furthermore, prior knowledge of the geometric formation of a vehicle path implies knowledge of its trajectory smooth analytical expression; it implies a real time positioning smoothing, in other words.In order to verify generalisation performance potentialities for the learning algorithm, a point classification application was implemented using real-world navigational data collected along a 120 km suburban railway track connecting Athens with Kiato at the north of Peloponnesus (Greece). According to the experimental results, two classifiers of different training length have both carried out a fine generalisation performance along unknown routes, recognising the linear pattern on any of the unseen point fixes at a level of 95%. Moreover, the implementation of the “real time positioning smoothing” in conditions of bad perception/outage of satellite signal delivered some significant corrections with regard to a post processed reference trajectory.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2016
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Frangos, Kyriakos
- Φράγκος, Κυριάκος
Subjects
dc:subject × 14- Πλοήγηση
- Μηχανική μάθηση
- Αναγνώριση προτύπων
- Βελτιστοποίηση
- Navigation
- Machine learning
- Pattern recognition
- Optimisation
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Πολιτικού Μηχανικού
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Civil Engineering
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
dc:rights- Statement dc:rights
-
- BY
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/39181