An Evaluation of New Global Appearance Descriptor Techniques for

Visual Localization in Mobile Robots under Changing Lighting

Conditions

Vicente Rom

, Luis Pay

, Sergio Cebollada

, Adri

an Peidr

and

Oscar Reinoso

Engineering Systems and Automation Department, Miguel Hernandez University, Elche (Alicante), Spain

Keywords:

Localization, Mobile Robots, Global Appearance Descriptors, Omnidirectional Images.

Abstract:

Autonomous robots should be able to carry out localization and map creation in highly heterogeneous zones.

In this work, global appearance descriptors are tested to perform the localization task. It focuses on the use of

an omnidirectional vision sensor as unique source of information and global appearance to describe the visual

information. Global-appearance techniques consist in obtaining a unique vector that describes globally the

image. The main objective of this work is to propose and test new alternatives to build and to handle global

descriptors. In previous experiments the images have been processed without considering the spatial distribu-

tion of the information. In contrast, in this work, the main approach is that relevant information will be in the

central rows. For this reason central rows information is given a higher weight comparing to other zones of the

image. The results show that this consideration can be an interesting presumption to take into account. The

experiments are carried out with real images that have been taken in two different heterogeneous environments

where simultaneously humans and robots work together. For this reason, variations of the lighting conditions,

people who occlude the scene and changes on the furniture may appear.

1 INTRODUCTION

In the last decade, the use of visual information has

extended to improve the autonomy of mobile robots in

many applications. The presence of mobile robots in

different environments has increased, and in order to

cope with more challenging situations their abilities

in perception and interpretation have improved. To

be robustly autonomous in extended, heterogeneous

and changing environments, the mobile robot has a

twofold task. First, a mapping task in which it has to

be able to navigate around the initialy unknown en-

vironment while creating a map. Second, it has to

perform localization task trying to estimate its posi-

tion and orientation in the environment. Among vi-

sion sensors, omnidirectional cameras are an interest-

ing option to carry out these tasks due to their ﬁeld

of view of 360

around the camera axis (Sturm et al.,

2011) and (Pay

a et al., 2017).

https://orcid.org/0000-0002-3706-8725

https://orcid.org/0000-0002-3045-4316

https://orcid.org/0000-0003-4047-3841

https://orcid.org/0000-0002-4565-496X

https://orcid.org/0000-0002-1065-8944

Due to the fact that images contain a big amount

of data, it is required to extract from them relevant

information. Nowadays local appearance descriptors

are well-known and extensively used. These meth-

ods describe speciﬁc points or local zones in the im-

age. Among these descriptors SIFT (Lowe, 2004)

and SURF (Bay et al., 2008) are the most known and

used. Murillo et al. (Murillo et al., 2007) solved

a mobile robots navigation problem using local de-

scriptors, Gil et al. (Gil et al., 2011) and Valiente et

al. (Valiente Garc

ıa et al., 2012) worked with local

appearance descriptors and omnidirectional cameras.

Relatively good results in navigation have been ob-

tained using local appearance descriptors. Global ap-

pearance descriptors are an alternative method to ex-

tract characteristic information from images and use

this information for mapping and localization.

Global-appearance description methods describe

the image globally obtaining a unique vector per im-

age, which is expected to be more invariant against

global changes in the scene. In addition, as each im-

age is described with a unique vector, the mapping

and localization work is simpliﬁed to a pairwise com-

parison between vectors. Over the past few years

some global-appearance descriptors have been stud-

Román, V., Payá, L., Cebollada, S., Peidró, A. and Reinoso, Ó.

An Evaluation of New Global Appearance Descriptor Techniques for Visual Localization in Mobile Robots under Changing Lighting Conditions.

DOI: 10.5220/0009595603770384

In Proceedings of the 17th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2020), pages 377-384

ISBN: 978-989-758-442-8

377

ied. Gist, introduced by Oliva and Torralba (Oliva and

Torralba, 2001), is one of the most extended descrip-

tors, and it has been tested in outdoor environments

for example by Zhou et al. (Zhou et al., 2018) to solve

the localization through matching the robot’s current

view with the best keyframe in the database. Other

option is the Histogram of Oriented Gradients (HOG),

HOG is used in (Pay

a et al., 2018) to solve hierar-

chical mapping and localization tasks. In addition,

there are other important techniques to obtain global

appearance descriptors based on mathematical trans-

formation, such as the Fourier Transform (Menegatti

et al., 2004) or the Radon Transform (Radon, 2005).

These alternatives have been used in works as Paya et

al. (Pay

a et al., 2009) to build maps or in Berenguer

et al. (Berenguer et al., 2019) where they used Radon

Transform to estimate the relative height of a mobile

robot. Moreover, during the last recent years, some

authors have used deep learning techniques to cre-

ate new global appearance descriptors. For example,

Xu et al. (Xu et al., 2019) proposed a CNN-based

descriptor to obtain the most probable robot position

and Cebollada et al. (Cebollada et al., 2019) perform

a comparison between analytic global-appearance de-

scriptors and CNN-based descriptors while solving a

mobile robot localization work. Finally, Rom

an et al.

(Rom

an et al., 2018) studied some of these global ap-

pearance methods in real environments to solve the

localization task under illumination changes.

As shown, global appearance descriptors are de-

ﬁned to be invariant against rotations in the ground

plane when omnidirectional images are used. Global-

appearance methods have summarised the informa-

tion from the panoramic images in horizontal blocks

or cells traditionally. But more recently, other ways

to build the descriptor have appeared, for instance

(Rom

an et al., 2019) where vertical cells are evaluated

with interesting results. The current work tries to go

one step beyond in the deﬁnition of global-appearance

descriptors, considering that usually, the most impor-

tant is condensed in the horizontal cells situated in the

middle of the panoramic image, because the visual in-

formation in the upper and bottom rows often corre-

sponds to the roof or sky and ﬂoor or terrain, which

are visually less signiﬁcative. For this reason a tech-

nique that increases the weight of the central rows is

studied. This work compares the classic formulation

with a new technique while testing them in a localiza-

tion framework.

2 GLOBAL APPEARANCE

DESCRIPTORS

In this section a review of the global appearance de-

scriptors used in the presented localization task is de-

scribed. The goal of these methods is to extract a

unique vector that globally describes the information

from an image. In this way, relevant information is

keep while reducing amount of memory. Global ap-

pearance descriptors have been used to perform robot

navigation tasks, for example, to solve the kidnapped

robot problem in indoor environments under different

conditions (Su et al., 2017) or to build hierarchical

maps through clustering algorithms (Cebollada et al.,

2019). To perform the localization task, HOG and

Gist descriptors are modiﬁed and used in this work.

In both cases the starting point is a panoramic image

i(x,y) ∈ R

×N

and after these methods each image is

reduced to a vector

d ∈ R

l×1

The ﬁrst step to build the descriptor is divide the

image in a set of cells. The descriptor size depends

on the number of these cells. The ﬁrst option stud-

ied in this work is the classic way, used in (Rom

et al., 2018) where the vector is built with uniformly

distributed and non-overlapped horizontal cells, ﬁg-

ure 1 shows how cells are distributed in this classi-

cal method. Taking into account the idea of giving

more importance to the central rows, a method where

the descriptors are built as traditionally but they are

weighted by a set of factors is evaluated. These fac-

tors are obtained from a Gaussian distribution centred

in the central row of the panoramic image, in such a

way that the information of the central cells is given

more importance than top or bottom cells. This idea

is outlined in ﬁgure 2.

Horizontal Cells Descriptor

Vertical Cells Descriptor

)

Figure 1: Classical approach to build the HOG global ap-

pearance descriptor of a panoramic image, by deﬁning a set

of horizontal cells.

Figure 2: Method to build the global-appearance descriptor

multiplying descriptor by a gaussian distribution.

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

378

2.1 Histogram of Oriented Gradients,

HOG

Histogram of Oriented Gradients was described by

Dalal and Triggs (Dalal and Triggs, 2005) and used

by Hofmeister et al. in small and controlled en-

vironments (Hofmeister et al., 2009). It basically

consists in calculating the gradient of the image and

obtaining magnitude and orientation of the gradient

in each pixel. After that, magnitudes and orienta-

tions are evaluated and the descriptor is build by

collecting together the information obtained in each

cell. The methods divide the panoramic images using

uniformly distributed and non-overlapped horizontal

cells. As they collect information from horizontals

rows and panoramic images are used, descriptors are

invariant to rotations of the robot in the ground plane

and they can be use in the localization tasks indepen-

dently on the orientation that the robot has at a spe-

ciﬁc time instant.

The descriptor size depends on diverse parame-

ters. This way, the vector length depends on the num-

ber of bins of the orientations histogram b and the

number of cells in which the image is divided k

HOG descriptor reduces a panoramic image into a

vector whose size is

d ∈ R

b·k

×1

2.2 Gist

This descriptor was initially proposed by Oliva et al.

(Oliva and Torralba, 2006) and it was developed by

Siagian et. al. (Siagian and Itti, 2009) testing its

performance in three different outdoor environments.

This method exposes the image to a speciﬁc number

of Gabor ﬁlters with different orientations in several

resolution levels. After that, the images are reduced

evaluating their mean intensity in different horizontal

cells.

In this case, the descriptor size will depend on the

number of orientations of Gabor ﬁlters m, the number

of cells in which the images are split k

and the num-

ber of different resolution models used r. During the

experiments this latter parameter r will be constant,

r=2. For that reasons the descriptor is a vector whose

size is

d ∈ R

2·m·k

×1

2.3 Main Parameters of the Descriptors

The aim of this work is to check the efﬁciency of the

descriptors in a localization task and to optimize the

main parameters in this new description technique.

Different parameters can be tuned and, as a conse-

quence, the vector that describes the image modiﬁes

its values and size. These parameters can be seen

Table 1: Parameters that impact on the location process.

Descriptor Parameters

HOG b⇒ number of bins per histogram.

⇒ number of horizontal cells.

Gist m⇒ number of Gabor ﬁlters.

⇒ number of horizontal blocks.

r⇒ different resolution models.

During these experiments r is constant, r = 2, the other parame-

ters take values between [8-32]

Figure 3: Gaussian distributions to obtain the coefﬁcients to

weight each cell, considering different values of sigma.

in table 1. Different values have been tested but, as

shown, these parameters deﬁne the descriptor size.

The larger the descriptor is, the more information it

contains but the slower the process will be.

The mainly purpose of this work is to evaluate

the new technique to build descriptors. As pointed

out before, the most important information is situ-

ated on the central rows in a panoramic image and

the top and bottom rows are visually less relevant be-

cause they contain the ceiling and the ﬂoor. For this

reason, the descriptors are built using the traditional

decomposition of the image in horizontal cells, but

each cell is weighted with an importance coefﬁcient.

The set of coefﬁcients is obtained from a Gaussian

distribution that gives more importance to the cen-

tral cells. Therefore, the information in these cells

contributes to the ﬁnal descriptor to a greater extent.

Gaussian distribution can be built with different de-

viation values (σ). Figure 3 shows different Gaus-

sian distributions used during experiments depending

on σ, σ = {50, 100,250,500, 750, 1000}. Values in a

gaussian distributions sum 1 and they are distributed

depending on σ.

An Evaluation of New Global Appearance Descriptor Techniques for Visual Localization in Mobile Robots under Changing Lighting

Conditions

379

Figure 4: Real-life changes in an heterogeneous environ-

ment.

3 DATABASE

The experiments have been carried out through the

use of the COLD dataset (Pronobis and Caputo,

2009) and INNOVA dataset (Amor

os et al., 2018).

COLD database offers three different indoor trajec-

tories taken in three buidings (Freiburg, Saarbr

ucken

and Ljubljana ) and INNOVA database offers also an-

other indoor laboratory trajectory the Miguel Hernan-

dez University, Spain. While the COLD database tra-

jectories were taken in different times of the day so

they offer the same environment in three different il-

lumination conditions, INNOVA trajectory was only

captured under cloudy conditions (during the light

hours but the sunlight does not considerably affect

the shots). The selected databases offer the hetero-

geneous and dynamic environment needed to test the

proposed global-appearance methods. The robot trav-

els along some laboratories where people is normally

working, so it has to deal with changes in the environ-

ment such as people walking or position of furniture

and objects and also problems like lighting changes

and occlusions. In the image 4 it is possible to see

some shots taken from the database where it is possi-

ble to observe these real operation conditions.

Among the different trajectories offered by the

datasets, three routes have been chosen as test datasets

to carry out the experiments. Route 1: Freiburg Part

A, Path 2, size 3 (Pronobis and Caputo, 2009), and

Route 2: INNOVA (Amor

os et al., 2018). In addition

each dataset has its own training database. Training

database covers the same parts of the environments

with an average distance between images around 0.02

m. Other trajectory speciﬁcations can be seen in ta-

ble 2, where it is possible to see the number of im-

ages and average distance between images. Within a

Table 2: Number of images and distance between consecu-

tive images in each route depending on the environment and

lighting conditions.

Trajectory Number Distance

Database of images between Images

Test Route 1 (Cloudy) 2778 0.0370 ± 0.0149 m

Test Route 1 (Night) 2896 0.0357 ± 0.0192 m

Test Route 1 (Sunny) 2231 0.0462 ± 0.0213 m

Training Route 1 (Cloudy) 556 0.1835 ± 0.0594 m

Test Route 2 (Cloudy) 1450 0.1212 ± 0.0410 m

Training Route 2 (Cloudy) 750 0.2397 ± 0.0629 m

selected route all the different speciﬁcations (cloudy,

night and sunny) cover approximately the same tra-

jectory, but they where taken in different moments.

Training routes were taken in a cloudy environment.

At the end, the route 1 covers approximately 103 m

and route 2 176 m.

4 EXPERIMENTS

4.1 Model of the Environment

As explained in the previous section, each database

offers a training trajectory which follows approxi-

mately the same route than the test ones. The dis-

tance between images in training dataset is around

0.20m and they were taken during the light hours, but

the sunlight does not considerably affect the images

(cloudy conditions).

Once the model is built with the descriptors of

the training images, the method to solve the local-

ization task consists in comparing the descriptor of

each test image with the descriptors in the model.

The program compares descriptors and calculates the

nearest neighbor by means of the correlation distance

(d(~a,

b) = 1 −

). In this expression:

~a ∈ R

lx1

and

b ∈ R

lx1

where: a

,i = 1, ..., l ;

= [a

− a,..., a

− a];a =

∑

·a

and

= [b

− b,..., b

− b];b =

∑

·b

When the nearest neighbor is calculated the geo-

metric distance between the capture point of the test

image (ground truth) and the capture point of the near-

est neighbor in the model is obtained, and the result

is the error. This geometrical distance can be calcu-

lated because COLD and INNOVA databases offers

the coordinates where each image had been taken but

the coordinates have been only used as ground truth

to check the error. The localization task is carried out

with pure visual information.

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

380

4.2 Position Estimation

Initially, the training routes are used to create the ref-

erence model. Afterwards, to study the robustness

of the global-appearance descriptors, the test images

are used to solve the localization problem. The local-

ization process evaluates which image in the training

model is the most similar to each test image. This

process has been carried out with the different light-

ing (cloudy, night and sunny). The error is calculated

as the geometric distance between the capture points

of both images. After repeating the process using dif-

ferent descriptors sizes, ﬁgures 5, 6 and 7 show the

average error (m) using the classical method, with no

weighting of the rows. Each ﬁgure shows the aver-

age localization error obtained after considering all

the test images of a trajectory with speciﬁc lighting

conditions.

64 128 256 512 1024 2048

Descriptor size

0.045

0.05

0.055

0.06

0.065

0.07

0.075

0.08

Error [m]

Cloudy

HOG Freiburg

Gist Freiburg

HOG Innova

Gist Innova

Figure 5: Localization error (m) using cloudy dataset versus

descriptor size.

Both, HOG and Gist descriptor provide their best

results when they work with middle size descriptors.

Gist performs better under cloudy and night condi-

tions, with errors between 0.0455 m and 0.2266 m

respectively while HOG outputs an error of 0.0458 m

with cloudy conditions and 0.2337 m with night envi-

ronment. The sunny conditions have the most nega-

tive effect on the localization process, with an error of

1.903 m with Gist and 1.8675 m using HOG. On the

other hand with INNOVA database only cloudy test

images are available. Results are really similar and

the lowest errors are 0.063 m using HOG descriptor

and 0.0618 m using Gist.

The results of the proposed method, which in-

cludes the weighting coefﬁcients, are shown in the

next lines. The localization process is the same as

shown before but now the information in the descrip-

tors is weighted by a set of coefﬁcients obtained from

64 128 256 512 1024 2048

Descriptor size

0.22

0.24

0.26

0.28

0.3

0.32

0.34

0.36

Error [m]

Night

HOG Freiburg

Gist Freiburg

Figure 6: Localization error (m) using night dataset versus

descriptor size.

64 128 256 512 1024 2048

Descriptor size

1.8

1.9

2.1

2.2

2.3

2.4

2.5

2.6

Error [m]

Sunny

HOG Freiburg

Gist Freiburg

Figure 7: Localization error (m) using sunny dataset versus

descriptor size.

a Gaussian distribution. Sigma (σ) is varied in order

to evaluate the inﬂuence of this parameter. Figures 8,

9 and 10 show the average localization error (m) us-

ing the new proposed method. As before each ﬁgure

corresponds with a speciﬁc lighting conditions of the

test images condition (cloudy, night and sunny).

Better results are obtained using middle-high σ

values. On Freiburg route and using HOG descriptor

and this technique an error of 0.04598 m is obtained

when σ=500 and cloudy environment, 0.1915 m when

σ=500 and night environment and 0.7505 m when

σ=250 and sunny environment. With Gist descrip-

tor the errors obtained respectively were 0.04538 m

0.2040 m 1.5687 m all of them obtained with σ=250.

On the other hand the lowest error in the INNOVA

route are obtained using σ=750. The minimum error

is 0.0619 m using HOG descriptor and 0.0637 m us-

ing Gist descriptor. These experiments clearly show

An Evaluation of New Global Appearance Descriptor Techniques for Visual Localization in Mobile Robots under Changing Lighting

Conditions

381

50 100 250 500 750 1000

Sigma

0.045

0.05

0.055

0.06

0.065

0.07

0.075

Error [m]

Cloudy

HOG Freiburg

Gist Freiburg

HOG Innova

Gist Innova

Figure 8: Localization error (m) using cloudy dataset versus

sigma in the Gaussian distribution.

50 100 250 500 750 1000

Sigma

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0.55

Error [m]

Night

HOG Freiburg

Gist Freiburg

Figure 9: Localization error (m) using night dataset versus

sigma in the Gaussian distribution.

50 100 250 500 750 1000

Sigma

0.5

1.5

2.5

3.5

Error [m]

Sunny

HOG Freiburg

Gist Freiburg

Figure 10: Localization error (m) using sunny dataset ver-

sus sigma in the Gaussian distribution.

that, lower errors are obtained using this new method

where the descriptors are multiplied by a gaussian dis-

tribution. Best results are obtained with descriptor

size 256 or 512 and taking into account the gaussian

distribution, it is better to multiplied the image with

a distribution σ=250 or σ=500. It can be observed

on ﬁgures 11 and 12. They show the best result in

each conﬁguration and it is possible to observe that

using the proposed method the error is lower. The

improvement is specially important when the test im-

ages the images are taken from the sunny environment

database. As seen, localisation error decreases using

the proposed method, especially with HOG descrip-

tor. The methods improve the task but to use it prop-

erly we have to supposed that the image is well dis-

tributed; there is the same quantity of cell and ﬂoor,

they are not important on the scene and the camera

is moved parallel to the ground ﬂoor. If these condi-

tions are suit the proposed method should improved

the localization task results.

HOG

Cloudy Freiburg Night Freiburg Sunny Freiburg Cloudy Innova

Environment

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Error [m]

Classical Method

Proposed Method

Figure 11: Comparison between classical and proposed

methods vs environment while using HOG descriptor.

Gist

Cloudy Freiburg Night Freiburg Sunny Freiburg Cloudy Innova

Environment

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Error [m]

Classical Method

Proposed Method

Figure 12: Comparison between classical and proposed

methods vs environment while using Gist descriptor.

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

382

Table 3: Calculation time (s) of the localization process in

the classical method.

Time to build Time

Descriptor Size Descriptor (s) Position (s)

HOG 64 0.0057 0.0836

128 0.0070 0.0844

256 0.0081 0.0858

512 0.0106 0.0895

1024 0.0132 0.1255

Gist 128 0.0295 0.0839

256 0.0458 0.0854

512 0.0643 0.0896

1024 0.1041 0.1252

2048 0.1478 0.1396

Table 4: Calculation time (s) of the localization process in

the new proposed method.

Time to build Time

Descriptor Size Descriptor (s) Position (s)

HOG 64 0.0060 0.0836

128 0.0070 0.0844

256 0.0082 0.0858

512 0.0106 0.0895

1024 0.0134 0.1255

Gist 128 0.0297 0.0839

256 0.0476 0.0854

512 0.0692 0.0896

1024 0.1054 0.1252

2048 0.1478 0.1396

4.3 Computational Cost

A low error is an important characteristic to take into

account when choosing a descriptor. But the com-

putational cost is also an important issue to consider.

For that reason, the necessary time to run the algo-

rithms has also been studied. Table 3 using the classi-

cal method and table 4 using the proposed one show

the time used to build the descriptor and to estimate

the position. The data are given in seconds. As ex-

plained in section 2, the results depend on descriptors’

size and this one depends on the parameters. Taking

that into account, HOG descriptor size is

d ∈ R

b·k

×1

in the same way Gist descriptor size is

d ∈ R

r·m·k

×1

The experiments have been carried out with a CPU

8-Core Intel Xeon E5

 at 3GHz and using the math-

ematical tool Matlab

 . These time results are not

absolute, they depend of the computer which runs the

process. But they are comparable because all the cal-

culations have been done with the same machine.

The lower the parameters are, the shorter the de-

scriptor is and for that reason the runtime also is

lower. If table 3 and table 4 are compared, it is possi-

ble to observe that the time used to build the descrip-

tor is almost the same.

Finally, it is possible to observe that the process is

quicker using HOG descriptor that when Gist is used,

especially in the time used to build the descriptor.

5 CONCLUSIONS

The present work studies a new way to use global-

appearance descriptors. The new method is compared

with the traditional one in a localization task. The

study has been made in real scenarios which are spe-

cially challenging due to lighting conditions and hu-

man activity. Using only visual information, global-

appearance descriptors and the new proposed method

described throughout this paper are studied. Once

the images are described the performance of these

descriptors in a localization framework is compared.

Both, geometric localization error and the computa-

tional cost of the process have been studied and the

parameters have been optimised.

First, about the traditional method, while Gist and

HOG offer relatively good results in cloudy and night

environments, the sunny conditions result more chal-

lenging, and HOG presents comparatively better re-

sults than Gist in this case. Second, about the pro-

posed method, the experiments show that it presents

substantial improvements, especially in sunny envi-

ronment. About calculation times, the new method

runs as quick as the traditional one. Observing the re-

sults the proposed method improves the localization

task decreasing the error. Results obtained with HOG

descriptor are better than the results using Gist, even

though both minimise the localization error. As the

run time is practically the same using both methods,

using the proposed method where the image is multi-

plied by a gaussian vector may be a proper way to ob-

tain better results in localization tasks where the base

map has been built in an heterogeneous environment

where changing lighting conditions and human activ-

ity can take part.

This work can be the ﬁrst step to build more suit-

able description solutions in navigation tasks, spe-

cially when omnidirectional or panoramic images in-

clude a lot of information from the ceiling and/or the

ﬂoor. The studied alternative may complement other

classical description methods in order to achieve a ro-

bust localization. Multiply the descriptor by a gaus-

sian distribution should not be considered as a unique

solution but they can contribute towards obtaining re-

sults. Future works can include a study of a more ro-

bust solution, combining for example the new studied

techniques with other measurements or techniques.

An Evaluation of New Global Appearance Descriptor Techniques for Visual Localization in Mobile Robots under Changing Lighting

Conditions

383

ACKNOWLEDGEMENTS

This work has been supported by the General-

itat Valenciana and the FSE through the grant

ACIF/2018/224, by the Spanish government through

the project DPI 2016-78361-R (AEI/FEDER, UE):

“Creaci

on de mapas mediante m

etodos de apariencia

visual para la navegaci

on de robots.” and by General-

itat Valenciana through the project AICO/2019/031:

“Creaci

on de modelos jer

arquicos y localizaci

on ro-

busta de robots m

oviles en entornos sociales”

REFERENCES

Amor

os, F., Pay

a, L., Mar

ın, J. M., and Reinoso, O. (2018).

Trajectory estimation and optimization through loop

closure detection, using omnidirectional imaging and

global-appearance descriptors. Expert Systems with

Applications, 102:273–290.

Bay, H., Ess, A., Tuytelaars, T., and Van Gool, L. (2008).

Speeded-up robust features (surf). Computer vision

and image understanding, 110(3):346–359.

Berenguer, Y., Pay

a, L., Valiente, D., Peidr

o, A., and

Reinoso, O. (2019). Relative altitude estimation us-

ing omnidirectional imaging and holistic descriptors.

Remote Sensing, 11(3):323.

Cebollada, S., Pay

a, L., Valiente, D., Jiang, X., and

Reinoso, O. (2019). An evaluation between global

appearance descriptors based on analytic methods

and deep learning techniques for localization in au-

tonomous mobile robots.

Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-

dients for human detection. In 2005 IEEE Computer

Society Conference on Computer Vision and Pattern

Recognition (CVPR’05), volume 1, pages 886–893

vol. 1.

Gil, A., Valiente, D., Reinoso,

O., Fern

andez, L., and

Mar

ın, J. M. (2011). Building visual maps with a

single omnidirectional camera. In ICINCO (2), pages

145–154.

Hofmeister, M., Liebsch, M., and Zell, A. (2009). Vi-

sual self-localization for small mobile robots with

weighted gradient orientation histograms. In 40th In-

ternational Symposium on Robotics (ISR), pages 87–

91. Barcelona.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International journal of computer

vision, 60(2):91–110.

Menegatti, E., Maeda, T., and Ishiguro, H. (2004). Image-

based memory for robot navigation using properties

of omnidirectional images. Robotics and Autonomous

Systems, 47(4):251 – 267.

Murillo, A. C., Guerrero, J. J., and Sagues, C. (2007). Surf

features for efﬁcient robot localization with omnidi-

rectional images. In Robotics and Automation, 2007

IEEE International Conference on, pages 3901–3907.

IEEE.

Oliva, A. and Torralba, A. (2001). Modeling the shape

of the scene: A holistic representation of the spatial

envelope. International journal of computer vision,

42(3):145–175.

Oliva, A. and Torralba, A. (2006). Building the gist of a

scene: The role of global image features in recogni-

tion. Progress in brain research, 155:23–36.

Pay

a, L., Fern

andez, L., Reinoso,

O., Gil, A., and

Ubeda, D. (2009). Appearance-based dense maps

creation-comparison of compression techniques with

panoramic images. In ICINCO-RA, pages 250–255.

Pay

a, L., Gil, A., and Reinoso, O. (2017). A state-of-the-art

review on mapping and localization of mobile robots

using omnidirectional vision sensors. Journal of Sen-

sors, 2017.

Pay

a, L., Peidr

o, A., Amor

os, F., Valiente, D., and Reinoso,

O. (2018). Modeling environments hierarchically with

omnidirectional imaging and global-appearance de-

scriptors. Remote Sensing, 10(4):522.

Pronobis, A. and Caputo, B. (2009). COLD: COsy Lo-

calization Database. The International Journal of

Robotics Research (IJRR), 28(5):588–594.

Radon, J. (2005). 1.1

uber die bestimmung von funktio-

nen durch ihre integralwerte l

angs gewisser mannig-

faltigkeiten. Classic papers in modern diagnostic ra-

diology, 5:21.

Rom

an, V., Pay

a, L., Flores, M., Cebollada, S., and

Reinoso,

O. (2019). Performance of new global ap-

pearance description methods in localization of mo-

bile robots. In Iberian Robotics conference, pages

351–363. Springer.

Rom

an, V., Pay

a, L., and Reinoso,

O. (2018). Evaluating

the robustness of global appearance descriptors in a

visual localization task, under changing lighting con-

ditions. In ICINCO-RA, pages 258–265.

Siagian, C. and Itti, L. (2009). Biologically inspired mo-

bile robot vision localization. IEEE Transactions on

Robotics, 25(4):861–873.

Sturm, P., Ramalingam, S., Tardif, J.-P., Gasparini, S., Bar-

reto, J., et al. (2011). Camera models and fundamental

concepts used in geometric computer vision. Founda-

tions and Trends

 in Computer Graphics and Vision,

6(1–2):1–183.

Su, Z., Zhou, X., Cheng, T., Zhang, H., Xu, B., and Chen,

W. (2017). Global localization of a mobile robot us-

ing lidar and visual features. In 2017 IEEE Interna-

tional Conference on Robotics and Biomimetics (RO-

BIO), pages 2377–2383. IEEE.

Valiente Garc

ıa, D., Fern

andez Rojo, L., Gil Aparicio, A.,

Pay

a Castell

o, L., and Reinoso Garc

ıa, O. (2012). Vi-

sual odometry through appearance-and feature-based

method with omnidirectional images. Journal of

Robotics, 2012.

Xu, S., Chou, W., and Dong, H. (2019). A robust indoor lo-

calization system integrating visual localization aided

by cnn-based image retrieval with monte carlo local-

ization. Sensors, 19(2):249.

Zhou, X., Su, Z., Huang, D., Zhang, H., Cheng, T., and Wu,

J. (2018). Robust global localization by using global

visual features and range ﬁnders data. In 2018 IEEE

International Conference on Robotics and Biomimet-

ics (ROBIO), pages 218–223. IEEE.

ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics

384