Feature reprojection as image alignment score in GNSS-free UAV localization
Roznovjak, Martin (2022)
2022
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2022113025322
https://urn.fi/URN:NBN:fi:amk-2022113025322
Tiivistelmä
Accurate absolute localization, commonly achieved with the help of Global Navigation Satellite Systems (GNSS), is critically important for unmanned aerial vehicle (UAV) navigation. However, GNSS can become unavailable or disturbed in certain environments. Therefore, different alternative solutions have been explored in the literature. A class of alternative approaches to absolute localization takes advantage of high-resolution satellite or aerial imagery and onboard video equipment.
The primary goal of this study was to assess the feasibility and characteristics of a novel solution to the visual self-localization of aerial video footage without the use of GNSS. The goal was achieved by developing, implementing, and evaluating a proof-of-concept solution and comparing it with the performance of a benchmark solution. The study was carried out for Huld Ltd. which has been developing a related product.
Both, the proposed and the benchmark solutions are Monte Carlo localization methods and use no other sensory information than video input and orthographic reference imagery (map). For absolute pose estimation, the proposed localization solution reprojects features from reference imagery to a query image. For comparison, the benchmark localization solution compares the query image to a corresponding part of the reference imagery using zero-mean normalized cross-correlation. The performance of the solutions was evaluated and compared on a diverse dataset of real aerial video footage.
The proposed localization solution achieved robust and consistent performance across the dataset and reached significantly lower localization error compared to the benchmark solution. The proposed method was found to be easily extendible and required no domain-specific engineering or tuning.
The primary goal of this study was to assess the feasibility and characteristics of a novel solution to the visual self-localization of aerial video footage without the use of GNSS. The goal was achieved by developing, implementing, and evaluating a proof-of-concept solution and comparing it with the performance of a benchmark solution. The study was carried out for Huld Ltd. which has been developing a related product.
Both, the proposed and the benchmark solutions are Monte Carlo localization methods and use no other sensory information than video input and orthographic reference imagery (map). For absolute pose estimation, the proposed localization solution reprojects features from reference imagery to a query image. For comparison, the benchmark localization solution compares the query image to a corresponding part of the reference imagery using zero-mean normalized cross-correlation. The performance of the solutions was evaluated and compared on a diverse dataset of real aerial video footage.
The proposed localization solution achieved robust and consistent performance across the dataset and reached significantly lower localization error compared to the benchmark solution. The proposed method was found to be easily extendible and required no domain-specific engineering or tuning.