Vision-based GNSS-Free Localization for UA Vs in the Wild Marius-Mihail Gurguy Jorge Pe na Queraltay Tomi Westerlundy

2025-05-06 1 0 6.36MB 6 页 10玖币

侵权投诉

Vision-based GNSS-Free Localization for

UAVs in the Wild

Marius-Mihail Gurgu†, Jorge Pe˜

na Queralta†, Tomi Westerlund†

†Turku Intelligent Embedded and Robotic Systems (TIERS) Lab

University of Turku, Finland.

Emails: 1{mmgurg, jopequ, tovewe}@utu.ﬁ

Abstract—Considering the accelerated development of Un-

manned Aerial Vehicles (UAVs) applications in both industrial

and research scenarios, there is an increasing need for localizing

these aerial systems in non-urban environments, using GNSS-

Free, vision-based methods. Our paper proposes a vision-based

localization algorithm that utilizes deep features to compute

geographical coordinates of a UAV ﬂying in the wild. The method

is based on matching salient features of RGB photographs

captured by the drone camera and sections of a pre-built

map consisting of georeferenced open-source satellite images.

Experimental results prove that vision-based localization has

comparable accuracy with traditional GNSS-based methods,

which serve as ground truth. Compared to state-of-the-art

Visual Odometry (VO) approaches, our solution is designed for

long-distance, high-altitude UAV ﬂights. Code and datasets are

available at https://github.com/TIERS/wildnav.

Index Terms—UAV; MAV; GNSS-Free; GNSS-denied;

Vision-based localization; Photogrammetry; Computer vision;

Perception-based localization; Visual Odometry

I. INTRODUCTION

Unmanned Aerial Vehicles (UAVs) are currently used in

a wide range of scenarios and commercial applications. They

already have reliable localization methods based on the Global

Navigation Satellite System (GNSS) [1], Visual-Inertial Simul-

taneous Localization and Mapping (VI-SLAM) [2], or Ultra-

wideband (UWB) systems [3]. The ﬁrst method is usually used

in outdoor environments, where it is relatively simple to use

a GNSS sensor to obtain accurate positioning data [4]. The

second and third methods prove their utility in environments

where GNSS signal is unreliable, for example inside buildings,

where localization systems such as UWB [5], [6] or VI-SLAM

are preferred [7].

Autonomy of aerial drones and their hardware capability

have signiﬁcantly improved in the last years [8], enabling

them to safely ﬂy autonomously over relatively long distances,

beyond visual line of sight (BVLOS). However, less attention

had been given to providing accurate positioning data during

long distance UAV ﬂights without using GNSS, which would

lead to more resilient solutions. This is an essential matter

——————————————————————————————

This research work is supported by the Academy of Finland’s AutoSOS

and AeroPolis projects (Grant No. 328755 and 348480) and by the Finnish

Ministry of Defence’s Scientiﬁc Advisory Board for Defence (MATINE)

project WildNav.

Fig. 1: Conceptual illustration of the proposed approach. Drone

images are compared to a set of reference satellite images and a

match is found based on deep features from Superglue [11].

since UAVs are mission-critical systems in need of a failsafe

mechanism that can be automatically engaged when GNSS

signal becomes unavailable due to various reasons, such as

jamming or spooﬁng [9].

At the same time, image processing capabilities using deep

neural networks have been constantly improving in the last

decade, enabling even real-time recognition of objects and

segmentation. This observation also holds for the processing of

satellite image data, where artiﬁcial intelligence (AI) models

are trained to differentiate between important objects such as

buildings and rivers [10].

Building on top of available open source technology and

algorithms [11], [12], this paper aims to provide a reliable

positioning mechanism for UAVs by only using RGB pho-

tographs provided by a camera mounted on the drone and

open-source satellite images of the ﬂight area. The method

is based on deep neural image segmentation that enables the

localization of the UAV by extracting recognizable features

of buildings, roads, rivers and forest edges. These features are

further matched to static satellite images using a robust method

based on a graph neural network model [11].

Fig. 1 shows the core idea of the proposed visual-based

localization method, where an UAV ﬂying at a high altitude

– provided that the ground surface below it has enough

landmarks with salient features – is able to match the cam-

era stream with onboard, georeferenced, open-source satellite

arXiv:2210.09727v1 [cs.RO] 18 Oct 2022

images.

The paper is organized in 6 sections as follows. Section II

provides a brief overview of related works on which the

current paper is based. Section III explains the motivation

and applicability of the implemented vision-based algorithm,

while Section IV presents the approach used for determining

the absolute geographical coordinates of a UAV. Section V dis-

cusses the experimental localization results. Finally, Section VI

concludes the work and outlines future research directions.

II. BACKGROUND

Previous studies directly related to our paper include seman-

tic segmentation based path planning [13], localization using

open source Google Earth (GE) aerial images [14] and pose

estimation with neural networks trained with georeferenced

satellite photographs [15]. In addition to these, an open source

image segmentation model originally used for building virtual

worlds [16] can prove its utility in providing useful input both

for localization and navigation purposes, as proposed in [13].

Additionally, a graph neural network that provides fea-

ture computation and matching for outdoor images, Super-

Glue [11], became part of the implementation in our proposed

localization algorithm due to its remarkable performance of

matching features in photographs which signiﬁcantly differ

in perspective and lighting conditions. The model proved

its efﬁcient application not only for matching features, but

also for the perspective transformations (namely homography)

used in computing geographical coordinates when running the

implemented vision-based localization algorithm. Before se-

lecting Superglue as a feature matcher between drone camera

photographs and satellite images, template matching [17] and

SIFT features [18] were also taken into consideration. The

latter two options were dropped due to their relatively high

computation time and inaccuracy, compared to [11].

Another remarkable paper is [12], which uses an advanced

template-based matching algorithm to compute the pose es-

timation of a UAV using only images. Because of insufﬁ-

cient computation resources at the time, no implementation

is provided on the onboard computer of a drone. Another

major difference, compared to our approach, is that the drone

navigates speciﬁcally urban environments, where computing

and matching visual features is much less of a challenge,

compared to natural, in the wild, ﬂying areas.

Currently, the state-of-the-art of navigation approaches in

a priori unknown environments are Simultaneous Localiza-

tion and Mapping (SLAM) algorithms, which have arguably

reached a high-degree of robustness and reliability in the past

decade [19]. An important building block of SLAM algo-

rithms are VO methods [20], which allow UAVs to accurately

determine their position while navigating new environments.

However, one major limiting factor of these approaches is

that they make the assumption the UAV is ﬂying at a low

enough altitude that an RGB monocular or stereo camera can

easily track position shifts in detected features from frame to

frame. Our work focuses on high-altitude ﬂight (120 meters), a

situation where a different approach for localization is needed,

as presented in Section III.

III. PROVIDING GNSS-FREE VISION-BASED

LOCALIZATION

Our main goal is developing a localization algorithm that

does not rely on GNSS for long-distance ﬂights, but only

on a monocular wide-angle camera. This kind of approach

proves its utility in situations where GNSS signal cannot be

reliably used and as a failsafe alternative that enables the drone

to reach its goal position or at least land safely in a pre-

established location. New commercial implementations such

as autonomous drone delivery could use the new localization

method to improve the reliability of navigation.

The core attribute of the project is the environment in which

vision-based localization is provided: in the wild, denoting

natural (non-urban) environments where artiﬁcial structures

such as buildings and roads are sparse. This characteristic

transforms what would be a trivial feature matching and ho-

mography computation problem inside a city into a challeng-

ing process, due to the difﬁculty of ﬁnding salient features in

natural environments. Nevertheless, the ﬁnal implementation

is able to provide accurate localization results using a neural

network for feature matching between drone photographs and

satellite images.

Another signiﬁcant feature of the implemented localization

algorithm is that the pre-mapping process does not require the

UAV to ﬂy. The map is built using exclusively open-source

satellite images, with the objective of enabling autonomous

localization in any area where the drone can ﬂy, assuming it

is legally and physically possible (due to harsh environmental

conditions).

IV. METHODOLOGY

Accurate feature matching of images is only useful if

the drone camera photographs can be precisely linked to

geographical coordinates. An important assumption of our

proposed localization method is that the ﬂight area of the UAV

is known a priori, so a map for that speciﬁc zone can be built

and uploaded to the onboard computer for ofﬂine use. The

map is composed of rectangular sections representing RGB

satellite images with an approximate resolution of 1400×1200

pixels, collected from GE. Each one of these sections are

collected from the same perspective, with the camera view

perpendicular to the ground surface and from an altitude that

offers a similar ﬁeld of view to a wide angle camera. One

example of these sections can be observed in Fig. 2, where

the transparent white rectangle represents the georeferenced

map tile. There is a linear relation (see Equations 1 and 2)

between the pixel coordinates of the image ﬁle and absolute

geographical coordinates.

A. Georeferencing map sections

Because the ﬂight area is generally too large to be repre-

sented as only one image ﬁle, the map is split into different

section, each one of them with two distinct geographical

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Vision-basedGNSS-FreeLocalizationforUAVsintheWildMarius-MihailGurguy,JorgePenaQueraltay,TomiWesterlundyyTurkuIntelligentEmbeddedandRoboticSystems(TIERS)LabUniversityofTurku,Finland.Emails:1fmmgurg,jopequ,toveweg@utu.AbstractConsideringtheaccelerateddevelopmentofUn-mannedAerialVehicles(UAVs)applic...

展开>> 收起<<

Vision-based GNSS-Free Localization for UA Vs in the Wild Marius-Mihail Gurguy Jorge Pe na Queraltay Tomi Westerlundy.pdf

共6页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Vision-based GNSS-Free Localization for UA Vs in the Wild Marius-Mihail Gurguy Jorge Pe na Queraltay Tomi Westerlundy

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: