ORB-Based Map Fusion with Position Transformation

for Enhanced Pairwise Connection

Lucas Alexandre Zick

1,2

, Dieisson Martinelli

1,2

, Andre Schneider de Oliveira

and Vivian Cremer Kalempa

Graduate Program in Electrical and Computer Engineering, Universidade Tecnol

ogica Federal do Paran

a (UTFPR),

Curitiba, Brazil

Department of Information Systems, Universidade do Estado de Santa Catarina (UDESC), S

ao Bento do Sul, Brazil

Keywords:

Map Fusion, Robot Localization, ORB Algorithm, Pairwise Map Alignment.

Abstract:

This paper discusses developing and evaluating a map fusion algorithm based on ORB (Oriented FAST and Ro-

tated BRIEF) feature matching, designed to improve the integration of robotic occupancy grids. The algorithm

effectively merges maps generated by multiple robots, accommodating map size and orientation variations. A

key aspect of its functionality is the ability to accurately position robots within the fused map, even when

the overlap between maps is minimal. Comprehensive testing demonstrated the algorithm’s effectiveness in

identifying correlations between different map pairs and aligning them accurately, as well as its capability to

assess the success of the merging process, distinguishing between successful merges and those with inaccu-

racies. The ﬁndings indicate that this approach signiﬁcantly enhances the capabilities of multi-robot systems,

improving navigation and operational efﬁciency in complex environments.

1 INTRODUCTION

The presence of multi-robot systems in robotics has

signiﬁcantly expanded the range of possible solutions,

primarily improving task execution speed (Gautam

and Mohan, 2012). Due to the growing ease of imple-

menting these technologies, multi-robot systems have

become increasingly common in academic research

and industrial settings (Arai et al., 2002; Va

ak and

Herich, 2023). This collaborative approach allows

tasks to be performed more quickly and efﬁciently

and facilitates the resolution of complex problems

that would be challenging for robots operating inde-

pendently.

In addition to efﬁciency, using multiple robots

also provides additional advantages, such as the abil-

ity to perform simultaneous tasks at different points

within an environment (Gautam and Mohan, 2012).

This is particularly useful in scenarios such as ex-

ploring unknown environments, monitoring vast areas

(such as space exploration), and conducting search

and rescue operations (Chakraa et al., 2023). With ad-

vancements in communication and sensing technolo-

gies, robots can share information in real time, en-

abling better coordination and a more accurate under-

standing of the environment in which they operate.

However, this growing reliance on multi-robot

systems brings speciﬁc challenges to the forefront,

particularly in map fusion and the integration of sen-

sory information (Stathoulopoulos et al., 2023). The

ability of different robots to merge their mapping data,

especially in situations where the maps may be dis-

connected or incomplete, is crucial for ensuring func-

tional autonomous navigation and operational effec-

tiveness (Mukhopadhyay et al., 2023). Therefore,

research into robust and scalable map fusion tech-

niques becomes essential for the success of multi-

robot systems in real-world applications (Ahmed Jalil

and Kasim Ibraheem, 2023).

Among the most promising solutions to address

these challenges is the ORB (Oriented FAST and

Rotated BRIEF) algorithm, which is widely used in

robotics and computer vision applications (Ahmed

et al., 2024). ORB is characterized by its fast and efﬁ-

cient detection of keypoints and its robustness in rec-

ognizing visual patterns, even in situations with vari-

ations in orientation or scale (Karami et al., 2017).

Its application in map fusion allows for the identi-

ﬁcation of correspondences between different maps

generated by robots of varying models in dynamic

environments, facilitating the combination of visual

data and the construction of a uniﬁed map. Addition-

Zick, L. A., Martinelli, D., de Oliveira, A. S. and Kalempa, V. C.

ORB-Based Map Fusion with Position Transformation for Enhanced Pairwise Connection.

DOI: 10.5220/0013670500003982

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 22nd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 2, pages 185-193

ISBN: 978-989-758-770-2; ISSN: 2184-2809

185

ally, ORB provides an efﬁcient solution for handling

large-scale maps, demonstrating good scalability re-

garding processing time and memory usage (Sabry

et al., 2024).

In this context, this work aims to develop an ap-

proach for map fusion based on ORB, focusing not

only on the combination of the maps themselves but

also on the precise transformation of the positions of

multiple robots within the ﬁnal map. By effectively

integrating the mapping data from two robots, it be-

comes possible to construct a uniﬁed map of the en-

vironment and accurately transform the robots’ po-

sitions within the resulting map, ensuring more efﬁ-

cient, safe, and coordinated navigation. Additionally,

the ORB algorithm utilized in this approach can as-

sess the likelihood of overlap between the provided

maps. This capability allows the system to identify

scenarios where the maps are disconnected or have

minimal overlap, avoiding unnecessary processing for

non-connected maps and optimizing data fusion only

when relevant. Thus, the proposed solution is robust

enough to handle maps of varying sizes and scenar-

ios where the robots lack complete environmental in-

formation, providing a more efﬁcient and adaptable

solution for multi-robot systems.

The contributions of this work include the devel-

opment of a scalable map fusion technique based on

the ORB algorithm, designed to handle partial overlap

or disconnection between maps, and implementing a

mechanism to correctly transform and align the posi-

tions of robots within the ﬁnal map. The experimen-

tal validation of the approach was conducted exclu-

sively in simulated scenarios, using randomly gener-

ated maps, which allowed for controlled testing of the

technique’s robustness and effectiveness. Although

the results demonstrate the potential of ORB-based

map fusion in virtual multi-robot systems, continu-

ous real-time application in physical robots has not

yet been explored. This represents the next step to

expand the research, validating the approach in real-

world and challenging environments, where the vari-

ations of the physical world can further demonstrate

the technique’s feasibility in practical situations.

The main contribution of this work is an integrated

framework for map fusion that addresses the practi-

cal challenge of aligning maps with partial or discon-

nected overlaps while ensuring the correct transfor-

mation of robot positions into the uniﬁed reference

frame. While foundational techniques like ORB are

well-established, our work focuses on their applica-

tion within a system that can reliably determine when

to merge maps and how to accurately place each robot

within the resulting shared environment. This ap-

proach is validated through a quantitative error met-

ric that enhances the system’s reliability by discarding

low-quality fusions. The experimental validation was

conducted exclusively in simulated scenarios, which

allowed for controlled testing of the technique’s ro-

bustness. Although these results are promising, real-

time application on physical robots has not yet been

explored and represents the next step to validate the

approach in challenging real-world environments.

The structure of the paper is as follows: Section

Proposed Strategy provides a detailed presentation of

the algorithm’s development. Section Experiments

and Validation introduces the experiments and valida-

tions conducted, along with the obtained results. Fi-

nally, Section Conclusion offers concluding remarks

and discusses potential future work.

2 RELATED WORK

In the ﬁeld of map fusion for multi-robot systems,

various approaches have been proposed to enhance

the integration of mapping data from multiple robots.

One notable work is by Dieisson Martinelli, who de-

veloped a method that requires standard references

within the mapped environments (Martinelli et al.,

2023). While Martinelli’s approach effectively facili-

tates the merging of maps when such references are

available, it presents limitations in scenarios where

robots operate in disconnected or unreferenced envi-

ronments. In contrast, the proposed ORB-based al-

gorithm does not necessitate standard references be-

tween the maps, allowing for greater ﬂexibility and

applicability across diverse environments.

Moreover, one of the key advantages of our ap-

proach is its ability to identify non-overlapping maps

rather than forcing a potentially inaccurate overlap.

This capability enhances the algorithm’s robustness

in dealing with disconnected environments and pre-

vents the introduction of erroneous data into the ﬁnal

fused map.

However, it is important to acknowledge that

while our method demonstrates advantages regard-

ing reference independence and accurate identiﬁca-

tion of map relationships, it has yet to be validated

in real-world scenarios. The reliance on simulated

environments raises questions about the algorithm’s

robustness and effectiveness in dynamic and unpre-

dictable settings typically encountered in practical ap-

plications.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

186

3 ORB-BASED MAP FUSION

This section discusses the proposed methodology to

address the challenges of map fusion and robot posi-

tion transformation using the ORB algorithm. The

proposed approach focuses on accurately detecting

overlapping areas between maps and efﬁciently merg-

ing them while ensuring the correct transformation of

the robots’ positions in the ﬁnal uniﬁed map. The fol-

lowing subsections outline the key components of our

method, including map overlap detection, the fusion

process, and position transformation.

3.1 Overview of the ORB Algorithm

The ORB (Oriented FAST and Rotated BRIEF) al-

gorithm is a robust feature detection and description

method widely used in computer vision. Developed

to overcome the limitations of traditional feature de-

tectors like SIFT and SURF, ORB is both computa-

tionally efﬁcient and invariant to rotations and scales,

making it ideal for real-time applications (Karami

et al., 2017; Yan et al., 2023).

The ﬁrst step in the ORB algorithm is key point

detection using the FAST (Features from Accelerated

Segment Test) method. FAST identiﬁes corners in an

image, which are potential key points. Each detected

key point is then oriented based on the intensity cen-

troid, ensuring that the features remain consistent un-

der rotation (Karami et al., 2017; Yan et al., 2023).

Figure 1: Keypoints detected by the ORB algorithm over-

layed on the original image.

Once the keypoints are detected, the ORB algo-

rithm describes these points using the BRIEF descrip-

tor, which compares the intensity of pixels around the

keypoint. This results in a binary string representing

the key point’s feature, enabling rapid matching be-

tween different images (Wu, 2023).

The beneﬁts of using ORB include its invariance

to rotation and scale, allowing features to be recog-

nized regardless of the image’s orientation or size.

Additionally, the algorithm offers high computational

efﬁciency compared to other methods, enabling real-

time operations. As illustrated in Table 1, ORB op-

erates signiﬁcantly faster than SIFT and SURF, with

execution times of just 0.02 seconds compared to 0.25

seconds for SIFT and 0.08 seconds for SURF, while

achieving a match rate that is close to its counterparts

(Karami et al., 2017).

Furthermore, the results in Table 2 highlight

ORB’s robustness against noise, achieving a match

rate of 54.48%, slightly higher than SIFT’s 53.8%

which demonstrates ORB’s ability to maintain perfor-

mance even in challenging conditions. Additionally,

ORB shows its effectiveness in varied scenarios, as

indicated in Table 3, where it maintains competitive

matching rates across different rotation angles. This

combination of speed and accuracy makes ORB an

ideal choice for applications requiring reliable perfor-

mance under diverse and complex conditions (Karami

et al., 2017; Cong, 2024).

Table 1: Results of comparing the image with its scaled

version. Reproduced from Karami et al. (2017). (Karami

et al., 2017)

Algorithm Time(s) Kpnts1 Kpnts2 Matches Match Rate (%)

SIFT 0.25 248 1210 232 31.8

SURF 0.08 162 581 136 36.6

ORB 0.02 261 471 181 49.5

Table 2: Results of the image matching by adding 30%

of salt and pepper noise. Reproduced from Karami et al.

(2017). (Karami et al., 2017)

Algorithm Time (s) Kpnts1 Kpnts2 Matches Match Rate (%)

SIFT 0.115 248 242 132 53.8

SURF 0.059 162 385 108 39.48

ORB 0.027 261 308 155 54.48

Table 3: Matching rate versus the rotation angle. Repro-

duced from Karami et al. (2017). (Karami et al., 2017)

Angle 0 45 90 135 180

SIFT 100% 65% 93% 67% 92%

SURF 99% 51% 99% 52% 96%

ORB 100% 46% 97% 46% 100%

In multirobot mapping scenarios, ORB facili-

tates effective map fusion by providing robust feature

matching between disconnected maps. This capabil-

ity of detecting and describing features in real time

can enhance the efﬁciency of collaborative robotic

systems (Sing

eis, 2024).

In conclusion, the choice of the ORB algorithm

is fundamental for the proposed map fusion strat-

egy, as it ensures accurate detection and description

of keypoints while maintaining computational efﬁ-

ciency. Furthermore, ORB excels in handling the dy-

namic nature of map sizes, essential in robotics appli-

cations where varied and complex environments may

be explored (Cong, 2024).

3.2 Map Generation and Patch

Extraction

In this work, an algorithm was developed for the dy-

namic and random generation of structured maps, en-

suring that each created map is unique. The function

ORB-Based Map Fusion with Position Transformation for Enhanced Pairwise Connection

187

generate structured map(size) generates a map

represented as a 2D matrix of dimension size × size,

where the value zero (0) represents free spaces and the

value one (1) represents obstacles. The default size of

the map is 550 × 550 pixels.

The map creation involves adding various geomet-

ric shapes, such as rectangles, circles, and ellipses,

using auxiliary functions that allow the inclusion of

varied elements in the environment. Below, we de-

scribe some of these functions and the associated cal-

culations:

• add rectangle(x, y, width, height,

angle): This function inserts rectangles into the

map, potentially applying a rotation according

to the speciﬁed angle. The coordinates of the

rectangle’s vertices are calculated from (x, y),

where x and y represent the position of the

top-left corner. The dimensions of the rectangle

are given by the parameters width and height.

The calculation of the rotated coordinates is done

using the rotation matrix:

R =



cos(θ) −sin(θ)

sin(θ) cos(θ)



where θ is the rotation angle.

• add circle(x, y, radius): For the insertion

of a circle, the function calculates the points that

satisfy the equation, where (x, y) are the coordi-

nates of the center of the circle and (X, Y ) are the

coordinates of any point within the circle:

(X − x)

+ (Y − y)

≤ radius

• add ellipse(x, y, r1, r2, angle): With

this function, ellipses can be added to the map.

The equation of the ellipse is given by:

(X − x)

(Y − y)

≤ 1

where (x, y) represents the center of the ellipse

and r1 and r2 are the radii of the ellipse along the

X and Y axes, respectively.

Randomness is a crucial aspect of the algorithm.

The function add large areas() creates a non-ﬁxed

number of large areas, where the size is randomly de-

termined within a variable range.

The function add corridors() generates a non-

ﬁxed number of corridors, where the width of the cor-

ridor is also a random value, and the length is equal

to the size of the map. The starting coordinates for

these corridors are selected randomly, ensuring a var-

ied distribution of obstacles.

The function extract patch(map,

patch size) is responsible for extracting

patches from the generated map. The patch

is randomly rotated, and the coordinates for

extraction are calculated using the function

get random patch coordinates(map size,

patch size), which ensures that the starting coor-

dinates (x, y) of the patch are within the map limits.

The coordinates are obtained through:

start

= random.randint(0, map size − patch size[0])

start

= random.randint(0, map size − patch size[1])

This ensures that the extracted patch does not ex-

ceed the dimensions of the map. An example of gen-

erated map and patches is presented in Figure 2

Figure 2: Example of generated map and patches.

The algorithm for generating structured maps and

extracting patches provides a diverse and dynamic

testing environment, allowing for practical and com-

prehensive experiments in map fusion and pattern

recognition.

3.3 Map Fusion Process

The map fusion process is crucial for integrating in-

formation from different sources, especially in sce-

narios where maps may have varying sizes and reso-

lutions (Martinelli et al., 2023). Fusion begins with

detecting and describing features in the maps using

algorithms such as ORB (Oriented FAST and Rotated

BRIEF). Once key points and their descriptors are ex-

tracted, the next step is to ﬁnd matches between the

descriptors of the two maps. Using BFMatcher, the

correspondences that indicate which features of one

map correspond to features of the other are identiﬁed.

An example can be seen in Figures 3 and 4.

Figure 3: Keypoints extracted from a pair of maps.

After identifying the correspondences, it is nec-

essary to calculate the transformation that aligns the

maps. This transformation may include transla-

tion and rotation and is obtained through the func-

tion cv2.estimateAffinePartial2D, which uses

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

188

Figure 4: Lines illustrating similarities on connections be-

tween the maps, represented side-by-side.

the RANSAC method to minimize the impact of erro-

neous correspondences. The resulting transformation

is described by:



1 0 dx

0 1 dy



where dx and dy are the translations along the x and y

axes, respectively. The rotation angle θ is calculated

from the parameters of the transformation matrix. For

example, if the obtained transformation matrix is:

M =



cos(θ) −sin(θ) dx

sin(θ) cos(θ) dy



the angle can be calculated as:

θ = arctan



M[1, 0]

M[0, 0]



Here, the notation M[i, j] refers to the element in

the i-th row and j-th column of matrix M, follow-

ing zero-based indexing conventions common in pro-

gramming languages such as Python.

Once the transformation is determined, the next

step is to apply the rotation to the second map

and overlay the two maps using the function

overlay maps. The merged map for the maps on the

Figure 3 can be seen in Figure 5

Figure 5: Final result of the map merging process.

Finally, considering the scalability of the process,

the fusion should be capable of handling maps of dif-

ferent sizes. For this, interpolation techniques and

scaling adjustments can be applied to ensure the infor-

mation is correctly combined, even when the maps are

captured at different resolutions or dimensions. This

modular approach makes map fusion scalable and ro-

bust, making it applicable to a wide range of scenar-

ios, from robotics applications to geospatial mapping.

3.4 Overlap Precision Detection

Before applying the transformation and overlaying

the maps, evaluating the accuracy of the fusion per-

formed is essential. For this, we calculate the mean

absolute error between the overlapping areas of the

maps, which provides a quantitative measure of the

quality of the fusion. The following formula deﬁnes

the error:

Error =

∑

i=1

[i] − I

[i]|

where I

and I

are the pixel values in the overlapping

areas, and N is the total number of pixels in the over-

lap region. A low mean error indicates that the maps

have been successfully fused and that the information

is aligned accurately.

This error metric is used purely for evaluating the

quality of the alignment after the transformation is es-

timated, not for optimization. The Mean Absolute Er-

ror was chosen for its straightforward interpretation in

occupancy grids, where it directly represents the av-

erage difference in pixel values, providing a clear and

computationally inexpensive measure of fusion qual-

ity.

Figure 6: Good align-

ment with a low error of

6.43 (mean pixel differ-

ence).

Figure 7: Minor mis-

alignment with an error

of 11.88.

Figure 8: Noticeable

misalignment with an

error of 20.56, nearing

the discard threshold.

Figure 9: Poor align-

ment with a high error

of 28.49, indicating a

failed fusion.

To better illustrate the impact of the error metric,

Figures 6, 7, 8, and 9 show examples of map over-

lays with different mean absolute error values. These

ﬁgures provide visual insight into how the increasing

error affects the accuracy of the fusion:

ORB-Based Map Fusion with Position Transformation for Enhanced Pairwise Connection

189

• In Figure 6, the error is relatively low (6.43), in-

dicating that the maps have been well aligned, and

the overlap area shows minimal discrepancies be-

tween the pixel values.

• As the error increases, as seen in Figure 7 (error

= 11.88), slight misalignments between the maps

begin to appear. While the fusion is still some-

what accurate, the overlapping areas exhibit mi-

nor differences, which remain nearly impercepti-

ble to the human eye.

• In Figure 8 (error = 20.56), the discrepancies be-

come more noticeable and start to affect the preci-

sion of the fusion signiﬁcantly. At this point, the

error reaches a value where the alignment is no

longer reliable, and the map should be considered

for discarding.

• Finally, in Figure 9 (error = 28.49), the misalign-

ment becomes more pronounced, with substantial

differences between the overlapping areas. From

this point onwards, the error reaches a threshold

where the maps must be discarded due to poor

alignment and unreliable fusion results.

Thus, the error remains nearly imperceptible up to

around 20. Beyond this threshold, the fusion quality

deteriorates signiﬁcantly. This provides a clear met-

ric for determining when to consider or discard maps

based on their mean absolute error.

In addition to the mean absolute error, feature de-

tection and the quality of the matches between the de-

scriptors are crucial for the success of the fusion. The

effectiveness of the matching algorithm, such as BF-

Matcher, and the precision in identifying features di-

rectly inﬂuence the quality of the calculated transfor-

mation. If a signiﬁcant number of erroneous matches

are identiﬁed, it is possible that the fusion result may

not be satisfactory, even if the mean absolute error is

low.

To enhance the robustness of the fusion process,

additional techniques are employed, such as cross-

validation of matches and the application of ﬁlters to

eliminate outliers. These methods ensure that only

the most reliable matches are included in the ﬁnal fu-

sion. This comprehensive approach improves the ac-

curacy of the map fusion and increases the system’s

resilience to noise and variations in the input data.

3.5 Robot Position Transformation

When merging two maps from different robots, it is

essential to accurately transform the robot positions to

the uniﬁed map (Martinelli et al., 2023). This process

ensures that each robot’s position is correctly placed

relative to the newly aligned map. In this work, the

transformation of robot positions is achieved by using

the same homography matrix used for map alignment.

Given a robot’s initial position (x

, y

) in its lo-

cal map, the transformed position in the uniﬁed map

′

, y

′

) is calculated by applying the homography ma-

trix:





′









0 0 1













This transformation ensures that the robot’s posi-

tion in the uniﬁed map reﬂects the correct orientation

and scale of the newly fused environment. An exam-

ple can be seen in Figure 10, with robots being repre-

sented by the colored dots.

Figure 10: Robots positions transformation.

Ensuring the accurate transformation of robot po-

sitions is critical for several reasons. First, it allows

for precise coordination between the robots within the

shared environment, enabling them to communicate

and make decisions based on their relative locations

(Wang et al., 2023). Furthermore, in navigation and

obstacle avoidance tasks, having correct position in-

formation is crucial to avoid collisions and plan efﬁ-

cient paths (Martinelli et al., 2023). Lastly, this pro-

cess guarantees that any subsequent actions or up-

dates in the environment (placing landmarks or inter-

acting with objects) are executed in the correct spatial

context, preserving the integrity of the merged map.

3.6 Computational Efﬁciency

The computational efﬁciency of the proposed map

merging algorithm is driven by its ability to optimize

the processing of large-scale maps. Despite process-

ing the entire map initially, the algorithm is designed

to reduce overall computation time by halting the pro-

cess when certain conditions indicate that the result-

ing map alignment would be suboptimal. This feature

allows the system to avoid unnecessary processing

when there is no clear beneﬁt to reﬁning poor matches

further.

One key factor contributing to this efﬁciency is

the algorithm’s ability to reject maps that do not meet

quality thresholds before completing a full merge. In

cases where the matching quality is insufﬁcient, the

algorithm terminates early, preventing further waste

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

190

of computational resources. This early termination

mechanism ensures that computational power is only

spent on map alignments, likely contributing to an ac-

curate ﬁnal map.

Additionally, by efﬁciently handling cases of poor

overlap or misalignment, the algorithm ensures that

high-quality maps are not replaced or corrupted by

low-conﬁdence merges. Figures 11 and 12 show ex-

amples of a successful merge versus a case where

early termination was beneﬁcial, visually emphasiz-

ing the effectiveness of the selective merging pro-

cess. This selective merging process improves reli-

ability and signiﬁcantly reduces the time spent on un-

necessary calculations, particularly valuable in real-

time operations or when working with extensive map

datasets.

Figure 11: Succesful merge, with an error of 5.47.

Figure 12: Unsuccesful merge, with an error of 33.78.

In summary, the algorithm’s efﬁciency lies in its

ability to streamline processing by avoiding the full

computation of maps unlikely to improve the over-

all result. This targeted approach reduces the compu-

tational load while maintaining the integrity of well-

matched sections, thus balancing speed and accuracy

in large-scale map fusion tasks.

4 EVALUATION AND ANALYSIS

Two distinct experiments were conducted to assess

the performance of the proposed map fusion algo-

rithm based on ORB. The ﬁrst experiment aims to

evaluate the algorithm’s ability to detect the quality

of the map merging, particularly focusing on its accu-

racy in identifying false positives. This test will also

inform future algorithm adaptations into a cyclical

system for continuous operation and real-time map

updates. The second experiment visually compares

the results of the automatic map fusion performed by

the algorithm with manually merged maps, providing

a qualitative analysis of its performance.

4.1 Comparison of Automatic and

Manual Merging

This experiment aims to evaluate the qualitative per-

formance of the proposed ORB-based map fusion al-

gorithm by comparing its results with those of manu-

ally merged maps. The goal is to visually assess how

well the algorithm aligns the maps regarding struc-

tural consistency, continuity, and overall accuracy.

For this comparison, several pairs of maps were

selected, each with varying levels of complexity, ro-

tation, and translation. The chosen maps included

features such as walls, corridors, and obstacles com-

monly found in robotic occupancy grids. Each map

pair was processed using two methods:

1. Automatic Fusion: The proposed algorithm was

applied to merge the maps without manual inter-

vention. The algorithm used ORB feature match-

ing to detect key points and estimate the transfor-

mation required for alignment.

2. Manual Fusion: A single human operator con-

ducted a manual fusion process in parallel. The

manual fusion involved identifying overlapping

regions between the two maps and aligning them

by visually adjusting their rotation and translation.

The results of both methods were displayed side

by side to facilitate a direct comparison. Key qualita-

tive aspects were analyzed visually, including:

• Alignment Precision: The visual degree to which

the structural elements of the maps, such as walls

and obstacles, appeared to be correctly aligned.

• Continuity of Features: The smoothness and

consistency of the merged map, particularly at

the boundaries between the original maps, are as-

sessed through visual inspection.

• Visual Accuracy: The merged map’s overall co-

herence, focusing on preserving critical map fea-

tures without distortion, evaluated based on visual

assessment.

Figures 13 and 14 show the visual outputs of the

automatic and manual fusion, respectively. As ob-

served in this example, the automatic fusion demon-

strated the ability to understand the connection be-

tween the maps even with limited overlapping ar-

eas, indicating its robustness in handling sparse data.

However, this particular example exhibited a slight

imprecision due to the lack of information. This com-

parison is one of several tests conducted using various

map pairs to evaluate the algorithm’s performance un-

der different scenarios.

ORB-Based Map Fusion with Position Transformation for Enhanced Pairwise Connection

191

Figure 13: Automatic fu-

sion.

Figure 14: Manual fu-

sion.

4.2 Robot Positioning Accuracy Test

In this section, we evaluate the capability of the pro-

posed map fusion algorithm to position robots within

the generated map accurately. The positioning accu-

racy is crucial for effective navigation and operation

of robotic systems in real-world environments. This

test assesses how well the algorithm can align the

merged maps with the actual positions of the robots

based on randomly assigned coordinates.

The experimental setup involved several pairs of

patches extracted from a larger map, with robot posi-

tions randomly assigned within each patch. The steps

taken during the testing process were as follows:

1. Patch Extraction: Randomly extract two patches

from a larger map, each containing a robot placed

at a random position within the patch.

2. Map Fusion: Apply the proposed algorithm to

merge the extracted patches into a single map.

3. Position Estimation: After merging, estimate the

positions of the robots within the fused map based

on the algorithm’s output.

4. Visual Assessment: Compare the estimated po-

sitions visually in the merged map to evaluate the

algorithm’s accuracy.

The results of the robot positioning accuracy test

are illustrated in Figures 15, 16, and 17. The ﬁrst two

ﬁgures display the individual patches with randomly

assigned robots. In contrast, the third ﬁgure shows the

ﬁnal merged map and the calculated positions of both

robots. The accuracy of the algorithm can be visually

assessed in the merged map. In all conducted tests,

the algorithm demonstrated exceptional performance

in accurately transforming the positions of the robots

within the merged maps.

Figure 15: Patch 1 with

robot in red.

Figure 16: Patch 2 with

robot in green.

Figure 17: Final merged map with estimated robot posi-

tions.

5 CONCLUSION

This study presented an algorithm for map fusion uti-

lizing ORB feature matching, speciﬁcally designed

to merge robotic occupancy grids. The algorithm

demonstrates signiﬁcant potential for improving the

reliability and efﬁciency of robotic navigation sys-

tems.

One of the key advantages of the proposed algo-

rithm is its capability to operate with maps of varying

sizes and orientations effectively. This ﬂexibility is

particularly beneﬁcial in multi-robot systems, where

different robots may generate maps under different

conditions. By accurately merging these maps, the

algorithm enables robots to collaborate and navigate

within a uniﬁed environment representation, enhanc-

ing their ability to perform coordinated tasks.

Moreover, the algorithm can estimate robot posi-

tions within the fused maps, even with limited over-

lapping areas. This positioning accuracy is crucial for

runtime decision-making and navigation in dynamic

environments.

The efﬁciency of the ORB feature matching

method further distinguishes this algorithm from con-

ventional approaches. Its computational speed allows

for real-time processing, making it suitable for dy-

namic applications where timely responses are essen-

tial.

Overall, the results indicate that the proposed al-

gorithm is a valuable tool for enhancing the capabili-

ties of autonomous robots. The successful integration

of maps is vital for enabling these systems to operate

effectively in complex environments.

All experimental data, including generated maps

and test results, can be found in a public repository

This repository is a resource for further exploring and

validating the proposed method in robotic mapping

and navigation tasks.

Future work on the proposed map fusion al-

gorithm could focus on several enhancements and

practical applications. One signiﬁcant improvement

would be to adapt the algorithm to handle occupancy

grids with values ranging from -1 to 255 rather than

https://github.com/LucasZick/map fusion

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

192

being limited to binary values of 0 and 1. This exten-

sion would allow the algorithm to incorporate more

nuanced information about the environment, such as

varying degrees of uncertainty or likelihood of obsta-

cles.

Additionally, it would be beneﬁcial to implement

the algorithm on real robotic systems operating in

real-world environments. Testing in practical settings

would provide valuable insights into the algorithm’s

performance under dynamic conditions and with ac-

tual sensor data. This approach could lead to further

reﬁnements and optimizations, enhancing the algo-

rithm’s robustness and applicability in diverse robotic

applications.

Exploring these avenues would contribute to de-

veloping more sophisticated multi-robot systems, ul-

timately improving their efﬁciency and effectiveness

in navigating complex environments.

ACKNOWLEDGEMENTS

The project is supported by the National Council for

Scientiﬁc and Technological Development (CNPq)

under grant number 407984/2022-4; the Fund for

Scientiﬁc and Technological Development (FNDCT);

the Ministry of Science, Technology and Innovations

(MCTI) of Brazil; Brazilian Federal Agency for Sup-

port and Evaluation of Graduate Education (CAPES);

the Araucaria Foundation; the General Superinten-

dence of Science, Technology and Higher Education

(SETI); and NAPI Robotics.

REFERENCES

Ahmed, M. F., Fr

emont, V., and Fantoni, I. (2024). Ac-

tive collaborative visual slam exploiting orb features.

arXiv preprint arXiv:2407.05453.

Ahmed Jalil, B. and Kasim Ibraheem, I. (2023). Multi-robot

slam using fast lidar odometry and mapping. Designs,

7(5):110.

Arai, T., Pagello, E., Parker, L. E., et al. (2002). Advances

in multi-robot systems. IEEE Transactions on robotics

and automation, 18(5):655–661.

Chakraa, H., Gu

erin, F., Leclercq, E., and Lefebvre, D.

(2023). Optimization techniques for multi-robot task

allocation problems: Review on the state-of-the-art.

Robotics and Autonomous Systems, page 104492.

Cong, Y. (2024). Image stitching technology for police

drones using an improved image registration method

incorporating orb algorithm. Informatica, 48(2).

Gautam, A. and Mohan, S. (2012). A review of research

in multi-robot systems. In 2012 IEEE 7th interna-

tional conference on industrial and information sys-

tems (ICIIS), pages 1–5. IEEE.

Karami, E., Prasad, S., and Shehata, M. (2017). Im-

age matching using sift, surf, brief and orb: per-

formance comparison for distorted images. arXiv

preprint arXiv:1710.02726.

Martinelli, D., Kalempa, V. C., and de Oliveira, A. S.

(2023). Map merge and accurate localization in

multi-robot systems in real environments. In Iberian

Robotics conference, pages 26–38. Springer.

Mukhopadhyay, S., Umari, H., and Koirala, K. (2023).

Multi-robot map exploration based on multiple

rapidly-exploring randomized trees. SN Computer

Science, 5(1):31.

Sabry, E. S., Elagooz, S., El-Samie, F. E. A., El-Bahnasawy,

N. A., and El-Banby, G. M. (2024). Sift and orb per-

formance assessment for object identiﬁcation in dif-

ferent test cases. Journal of Optics, 53(3):1695–1708.

Sing

eis, R. (2024). Check for performance analysis of

orb-slam in foggy environments rita sing

eis, sedat

dogru (), and lino marques institute of systems and

robotics, department of electrical and computer. In

Robot 2023: Sixth Iberian Robotics Conference: Ad-

vances in Robotics, Volume 1, volume 1, page 209.

Springer Nature.

Stathoulopoulos, N., Koval, A., Agha-mohammadi, A.-a.,

and Nikolakopoulos, G. (2023). Frame: Fast and ro-

bust autonomous 3d point cloud map-merging for ego-

centric multi-robot exploration. In 2023 IEEE interna-

tional conference on robotics and automation (ICRA),

pages 3483–3489. IEEE.

ak, J. and Herich, D. (2023). Map merging for multi-

robotic applications. In 2023 IEEE 21st World Sympo-

sium on Applied Machine Intelligence and Informatics

(SAMI), pages 000021–000026. IEEE.

Wang, M., Cong, M., Du, Y., Liu, D., and Tian, X. (2023).

Multi-robot raster map fusion without initial rela-

tive position. Robotic Intelligence and Automation,

43(5):498–508.

Wu, K. (2023). Creating panoramic images using orb fea-

ture detection and ransac-based image alignment. Ad-

vances in Computer and Communication, 4(4):220–

224.

Yan, H., Wang, J., and Zhang, P. (2023). Application of

optimized orb algorithm in design ar augmented real-

ity technology based on visualization. Mathematics,

11(6):1278.

ORB-Based Map Fusion with Position Transformation for Enhanced Pairwise Connection

193