Visual Simultaneous Localization and Mapping with Deep Learning
No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Author
Date
2020-12-14
Department
Major/Subject
Autonomous Systems
Mcode
ELEC3055
Degree programme
Master's Programme in ICT Innovation
Language
en
Pages
67+4
Series
Abstract
One of the biggest challenges of the automotive industry at the moment is the idea of autonomous vehicles and the huge amount of data that they require due to the main technology they use, Deep Learning. Often, collecting enough data is very expensive and time-consuming, causing the industry to start adopting technologies such as Scenario Cloning, where previously recorded sequences are used to digitally reconstruct the scenario. At its time, within this field, one of the most relevant tasks is Simultaneous Localization and Mapping. This thesis presents a series of improvements based on Deep Learning that can be introduced in current feature-based Visual Simultaneous Localization and Mapping systems to overcome some of the most recurrent problems, such as dealing with highly dynamic environments. The main focus of the thesis is to take an existing state-of-the-art Visual Simultaneous Localization and Mapping method and combine it with Deep Learning-based semantic segmentation. The resulting system successfully avoids placing features on dynamic objects and other regions that tend to decrease the performance of the system, thus improving substantially the overall performance on dynamic environments. Additionally, the system uses the information provided by the Deep Learning model to assign semantic information to each of the points forming the sparse map, resulting in a more complete tool and opening the door for new opportunities in tasks such as obstacle avoidance or planning.Description
Supervisor
Zhou, QuanThesis advisor
Matskin, MihhailHåkansson, Anne
Keywords
deep learning, simultaneous localization and mapping, semantic segmentation, visual slam