Institutional Repository of Vilnius University
Study of the adaptation of visual localization algorithms to different flight conditions /
Abstract
dc:descriptionIn recent years, the utilization of Unmanned Aerial Vehicles (UAVs) has been increasing rapidly. With the Global Navigation Satellite System's (GNSS's) interference technologies only increasing in prevalence, such as jamming, there emerges a need for drone localization in GNSS-denied environments. One of the most utilized localization techniques as of today is visual localization. Currently, there exists a wide variety of visual localization algorithms, however, only a handful can prove to be effective in outdoor situations. In this work Semi-Direct Visual Odometry (SVO) algorithm is compared against a Discriminatory Pearson Correlation based Particle Filter Localization (DPC-PFL) algorithm in terms of accuracy and consistency on a series of non-ideal flight scenarios, to which both of the algorithms will be adapted to, in a simulated environment. The aim of this work is to examine which of the two algorithms perform better under which non-ideal scenario and to determine if these visual localization implementations are a viable mitigation strategy for GNSS interference. The scenarios used include varying frame rates, altitudes and texture quality of the chosen terrain. During experimentation streams of images and ground truth values were collected, together with metadata regarding relative pose and drone rotation to form datasets. Afterwards, an exhaustive parameter search method was implemented for each of the algorithms and each dataset to retrieve optimized configurations and adapt the algorithms to said datasets. Finally, multiple flights were localized with the same optimized configuration of each algorithm and dataset to provide average mean error (accuracy) and standard deviation of mean error (consistency) values. Compiled results were then visualized and compared between SVO and DPC-PFL. Results showcased DPC-PFL algorithm's dominance over SVO in all non-ideal flight conditions, and it constantly scoring a significantly lower average mean error and standard deviation showcasing its high accuracy and consistency. Yet both algorithms were able to accurately estimate the pose and produced favourable results under all non-ideal flight conditions in a simulated environment.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Variakojis, Arnas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308339043&prefLang=en_US