1 Problem significance and
motivation
Nowadays, cameras are widely embedded into many
devices like phones, wearable devices, robots, and
cars and they are used in a large variety of applications
providing efficient tools for improving the perception
capabilities of these devices. In order to achieve this,
deep-learning approaches are introduced to improve
performance quality of traditional computer-vision
tasks. However, these approaches are well known
to perform well only under optimal conditions, i.e.,
with input images/videos without noise, with high
quality, etc. However, such conditions are obviously
not always met, especially when these cameras are
operating under real-world environmental conditions
(also called environmental noise), including rain,
dust, fog-haze, fire-smog, light reflections, limited
illumination conditions, etc.; under well-known
hardware camera limitations (i.e., sensor noise,
non-linearity, white balance, image artifacts); or
under security breaches (e.g., adversarial attacks).
Despite the aforementioned limitations, deep
learning is taking over as the single most important
and powerful method for object-recognition and clas-
sification tasks. Deep Neural Networks (DNNs), de-
spite the diversity in their architectures, all share a
common nature. Their performance is as good as the
data they are trained with. In this context noise and
distortions in the training data become important is-
sues since high-quality annotated data are not easy to
find for every specific training task. Current systems
mainly rely on three general approaches in addressing
noise in training data:
• Incorporating noise in the recognition task
by training standard DNNs for classifica-
tion/recognition [1], [2], [3], [4], [5], [6], [7] (of
the suitable kind for each cue) with a mixture
of noisy and clear inputs, so the network will
learn how to correctly map noisy inputs as well.
In this scenario, however, a huge amount of
high-quality annotated noisy data is required,
leading to a chicken - egg dilemma very common
in deep-learning tasks today.
• Generating noisy training data of high annota-
tion quality by applying extended augmentation
techniques [8], [9], [10], meaning producing syn-
thetic images with various kinds of noise and dis-
tractions, different views, angles etc., similar to
the ones expected in real-life scenarios. Even
though this is a valid approach, there is no obvi-
ous way for one to assess how realistic these syn-
thetic data would be or how its divergence from
reality affects the learning accuracy and perfor-
mance in real-life scenarios.
• Applying denoising [15],[14],[16],[17] as an ex-
plicit preprocessing step to the inputs of the
AI-Powered Disinfection Force: Harnessing 5G for Intelligent
Distributed Object Recognition and Sanitation Fleet
NIKOLAOS K. PAPADAKIS
Department of Machines, Intelligent and Distributed Systems
Infili Technologies S.A.
60, Kousidi str, Zografou 15772, Athens
GREECE
Abstract: With the increasing prevalence of embedded cameras in devices such as phones, robots, and vehicles,
there’s a growing need to enhance their perception capabilities through advanced deep learning techniques.
These systems, however, often struggle under suboptimal conditions like environmental noise. The paper
addresses these challenges by proposing a 5G-enabled solution that uses a fleet of unmanned ground vehicles
(UGVs) for coordinated disinfection tasks in hospital environments. Utilizing high-bandwidth, low-latency 5G
networks, these UGVs can share high-quality images for distributed object recognition tasks and execute precise
disinfection routines using onboard UV lamps. The proposed system includes an optimal formation algorithm
for UGVs to maintain effective positioning and coordination. By leveraging 5G connectivity, the fleet can
efficiently exchange sensor data and perform real-time, innovative object recognition and disinfection,
presenting a significant advancement in intelligent sanitation technology.
Key-Words: AI; Computer Vision; Intelligent Fleet; Distributed Object Recognition; Decentralized
Architectures; Edge Computing; 5G
Received: April 6, 2024. Revised: September 11, 2024. Accepted: October 2, 2024. Published: October 29, 2024.
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
239
Volume 6, 2024
recognition/classification task. The advantage of
this approach lies in decoupling noise removal
from the recognition task in a way that the for-
mer can be learned explicitly. Deep learning can
thus be used for denoising as well, but with mod-
ified architectures that favor this specific task.
The challenges of this approach lies primarily in
choosing the correct denoising architecture de-
pending on the task, as a one-size-fits-all denois-
ing architecture does not exist.
No matter the approach, the need of extensive
and high-quality training is thus the major limitation.
This is because manifestations of the same object
can take arbitrary forms in the wild, e.g., images
taken from different angles, with occlusions, lighting
conditions and with varying noise effects as was
discussed above. It is thus impossible for one to
train with images that cover all possible cases. If
only a single image is presented for recognition the
probability of misclassification cannot be measured
under these varying conditions. However, if more
images of the same object from different angles and
under different conditions are presented simultane-
ously for classification with different algorithms,
the probability of a correct classification increases
without the need of additional training.
It is this idea that will be pursued in the use-
case scenario presented herein. Based on the
above-mentioned observations a new technology is
proposed that will enable a fleet of unmanned ground
vehicles (UGVs) to perform object recognition and
coordinated fleet-formation tasks for the purpose
of disinfecting indoor and outdoor hospital areas
using onboard UV lamps in a 5G-enabled setting.
The availability of high bandwidth low latency
5G networks makes it possible to coordinate the
exchange of high-quality images among a fleet of
UGVs that collectively perform a recognition task
by running different classification algorithms in a
suitable distributed manner and apply predefined
disinfection actions on the recognized object. We
propose to demonstrate how a fleet of UGVs carrying
sensors, cameras and adjustable UV lamps in a
5G infrastructure setting, can run different onboard
recognition systems and perform disinfection actions.
By arranging specific formations between them, each
UGV can collect images from multiple angles of the
same object with the purpose to optimize recognition
and disinfection actions. 5G connectivity enables
not only the exchange of high-resolution images and
video between single UGVs, but also a distributed
UGV formation algorithm which can run in real
time. By presenting many simultaneous views of
the same object at the querying phase one facilitates
the classification algorithms to generalize better on
the unseen object and decide the correct disinfection
routine. At the same time 5G connectivity enables a
mixture of the above techniques to be unified under
a boosting framework [13] by utilizing the fleet
hardware collectively in real time.
Additionally, for enhanced coordination and sys-
tem reliability, a hierarchical structure is proposed,
where the swarm is composed of leader UGVs at the
top level, regional UGVs at the mid-level, and in-
dividual UGVs at the bottom layer. This structured
approach ensures efficient coordination and task exe-
cution, Also it enables effective emergency situation
handling through a novel weighted consensus mecha-
nism that is proposed through this work based on the
Byzantine Fault Tolerance algorithm [21].
2 5G use-case scenario and its
significance
Using 5G connectivity we propose a new, distributed
visual-surveillance technology for extracting action-
able information and perform specific predefined
tasks, more specifically, disinfection indoors/outdoor
hospital areas using UV light lamps installed on
UGV fleets. Focusing on visual (and, possibly, in-
frared) imaging sensors the system should offer iden-
tification, classification and analysis on visual data
from UGVs as well as stationary visual surveillance
sources and enable real-time, onboard decisions and
system-wide planning regarding route, speed, and dis-
infection tasks. The suggested scenario is a small fleet
of 10-15 UGVs deployed at a 5G-enabled hospital in-
stallation with the mission to perform visual recogni-
tion tasks and UV-light disinfection on demand, from
various points around one building sharing corridors
and across uncontrolled lanes or through busy park-
ing lots to another building on the installation with
speeds ranging from 3-25mph. The UGVs are aided
by stationary networked cameras that cover the area,
a priori visual background data (enabling background
subtraction) and knowledge of own location. The de-
mand for stationary cameras are not restrictive since
any UGV can play this role and in fact stationary
UGVs will model a fixed coordinate system from op-
timally chosen positions through which other UGVs
will tag their position by intra-fleet 5G communi-
cations. The roles between stationary and moving
UGVs can change depending on the conditions.
The system should offer the following capabilities:
1. Obstacle detection and Obstacle Avoidance
(ODOA).
2. Correct positioning and speed regulation with re-
spect to moving and stationary objects.
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
240
Volume 6, 2024
3. Co-ordinated and optimized system-wide re-
sponses across the fleet.
4. Data collection and communications.
5. Extraction of actionable information from the
sensor stream.
6. Disinfection in rooms, corridors and individual
objects using coordinated fleet movements and
the use of onboard UV lamps.
A high-bandwidth and low-latency 5G envi-
ronment enables powerful distributed algorithms
for fleet coordination and actionable information
extraction. In current systems, units performing the
recognition and sensor carriers themselves, such as
UGVs, have only limited availability (bandwidth
and delay constrained) of the data coming from
stationary and other UGV nodes while they typically
have full availability of the data flow in their own
sensors. This constraint is lifted in 5G-connectivity
scenarios. In the proposed system, therefore, each
UGV analyses data and extracts data significant for
object identification for the dual purpose of: a) to
locally provide ODOA, b) to use the 5G bandwidth
to share with other UGVs high-resolution images and
videos of its own angle of view. These two purposes
are highly intertwined because other UGVs receiving
extracted data become rather effective in extracting
their own information for further sharing.
In the present use case 5G low latency enables a
coordinated action to minimize intra-group disruption
but also to optimize coordinated responses in a deter-
ministic fashion (subject to mission-critical, time and
other constraints). The usefulness of the sensor data,
collectively acquired by the fleet, is directly related to
field parameters (angle, distance, relative speed, etc.),
and, thus, the loss or lack of formation directly un-
dermines the ability of combining sensor data from
different UGVs towards credible recognition. Im-
ages taken from the same angle do not contribute to
further collective analysis. It becomes then evident
that certain formations may greatly facilitate corre-
sponding fusing algorithms by providing diversified
sensor data. This way, significant gains in computa-
tional time and accuracy in recognition tasks can be
achieved. A controlled formation can also diversify
the response among the UGVs and provide the means
to bind responses in a meaningful manner following
deterministic mission-critical specifications. For in-
stance, UGVs might coordinate in sharing the disin-
fection task in a room.
2.0.1 The Optimal Formation Algorithm
The formation algorithm runs on each of the UGVs
and determines the movement of each of the UGVs
for the disinfection task. The UGVs try to keep a
certain formation by solving a point-correspondence
problem. Each UGV knows the positions of the
other UGVs, the position of the obstacles and other
objects in the environment in space-time, and the
formation to be kept during disinfection. Therefore,
the movement of the whole fleet to the next position
can be modeled as a geometric point-correspondence
problem where each UGV must move from its current
position to the desired position that will bring it in the
desired formation, covering the minimum distance
to the ultimate target for disinfection. If a UGV has
knowledge of the full formation then it can move
simultaneously to cover a new position – a benefit
of keeping a formation while moving. If the desired
position cannot be occupied by a specific UGV,
due to an obscuring object or another constraint,
the formation becomes approximate. However, the
UGVs can learn the new approximation. The UGVs
that are out of formation enter a following mode by
solving a different correspondence problem, namely
that of staying out of the way of the UGVs that are
in formation, but following them while waiting for a
chance to get in formation again and opportunistically
exchanging positions with other UGVs already in
the formation. Note that all UGVs solve the same
correspondence problem of moving in formation to
new positions in a distributed manner knowing their
unique ID but also what all the other UGVs will be
attempting, given a predefined disinfection algorithm.
The proposed distributed algorithm is modeled by
solving a shortest-path problem on a graph in space
time. The space is quantized as a 3D bin space. Each
bin is represented by its tag or the coordinates of its
center and represents a vertex of the graph. Each bin
represents a location where objects are allowed to ex-
ist. Transitions from one bin to another are allowed
explicitly by inserting edges connecting respective
bins. As objects come in and go out of the scene (e.g.,
other UGVs), bins can be either occupied or free; in
the former case all edges to this bin are erased; in the
latter case edges connecting the bin to other neigh-
boring unoccupied bins are inserted. This is a local
to the bin operation and can efficiently be performed
in a completely distributed fashion (bits in a binary
mask representing the graph). Each UGV runs the
same algorithm, and at any given time each UGV has
the following information, common to all UGVs (5G
guarantees common real-time information).
1. The space bin graph (defines occupied bins in
space, including other UGVs).
2. A certain formation the UGVs have to follow.
3. A protocol of synchronous movement related to
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
241
Volume 6, 2024
the current formation.
4. A target object to be disinfected and a predefined
algorithm of movement to do so.
Step 3 above requires a synchronous phase.
Alternatively, it can be implemented by circulat-
ing a token between UGVs in an order dictated
by the specific formation, this approach can be
implemented seamlessly on top of 5G networking
protocols necessary for the communications between
UGVs. This unification of movement, recognition
and disinfection task in the context of networking
protocols is the main innovation of the proposed
use case system particularly important in view of
the mutual optimization choices and seamless 5G
integration.
The UGV currently holding the token solves
the shortest path algorithm considering as occupied
bins the ones occupied by all the existing objects
in the environment together with their (immediate)
future positions (extrapolated by speed and direction
information derived from the predefined disinfection
algorithm of movement) and the UGVs that already
had the token and already moved to their new posi-
tions. In other words the order in which the token
is exchanged (a topological parameter of the current
formation) governs the UGVs’ order of movement
and is the temporal part of the proposed algorithm.
For a certain UGV there is a preference for the bins
closest to it that keep it in the task and in formation.
This preference is modeled by introducing weights
related to the edges connecting the bins. If the UGV
can move to the desired bin that keeps it in formation
the algorithm continues as above and the collective
movement remains normal and in formation.
Networking communications are also optimal
(maximizing the benefit of 5G infrastructure in terms
of bandwidth requirements for the ad hoc network
between UGVs and stationary nodes) and sensor data
are also optimally exchanged. If a certain UGV is
not able to move to the desired bin (according to
formation position) it drops to a following mode,
by moving at the end of the queue in the next token
run, or changes disinfection task by moving into
another area. This, in practice, means that at the next
run it will move last, trying to occupy some other
valid position. Its token order will move upwards
progressively as other UGVs lose formation and
become last in the token queue. When a ’following’
node succeeds in entering a valid formation bin, it
moves upwards in the token queue, in front of all
the UGVs that remain in following mode. Some
important observations:
• Each UGV solves an “all shortest paths” algo-
rithm according to Ref. [11] which is suitable for
real time application. It can thus predict move-
ments of other UGVs and thus optimize its own
movements during disinfection.
• The concept of “formation” is used only as a
“topological” concept, i.e., it dictates the order
of exchanging a token for movement and possi-
bly the networking protocol for communications
that is optimized by the certain formation and
5G infrastructure. Specific to the certain topo-
logical formation metric parameters (distance be-
tween the UGVs, closest distance permitted for
other objects, etc.), can be dynamically adapted
by modifying bin sizes in the proposed graph rep-
resentation in 3D space.
• Bin sizes, e.g., quantization/granularity of time
and space resolution can be dynamically adjusted
according to terrain/mission constraints; speed
regulation can also be modeled this way.
• The direction of onboard UGV cameras and UV
lamps can be dynamically directed to certain bins
resolving ambiguous graph-connectivity issues
and optimizing bin resolution parameters.
3 The progressive object detection
algorithm
According to the terrain modeling as a 3D bin grid,
the object-detection procedure initially will mark cer-
tain bins as ‘occupied’ when an object appears in the
range of detection. The contribution of the stationary
cameras (or ‘stat UGVs’) is important towards this
direction. Each stationary camera observes a certain
range of bins. A simple background removal can be
implemented with simple hardware at the stationary
camera by superimposing the current captured image
on a clean plate containing a static background. A bi-
nary quantization of the resulting image can produce
a binary mask marking the coordinates occupied by
the object. All such masks are transmitted to UGVs
whenever a significant change happens in the scene.
At the UGV side, binary masks are collected from
stationary camera broadcasts. A pair of orthogonal
views of the same scene (same bins), results in
extending the bin mask as a 3D representation. A
combination of several masks from various stationary
views of the same scene further refines the 3D bin
mask showing the occupied bins. This procedure
of assembling the 2D stationary masks to 3D bin
masks and the respective updates of the occupied bins
happens at the UGV side using the stationary data.
It is the object-detection phase where the interest
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
242
Volume 6, 2024
is restricted to the specific bins occupied by the
objects. A UGV can now use its onboard camera to
focus on its respective bins and take a high-definition
image/video of the respective object. These high-
resolution (hi-res) images containing the object of
interest can now be used locally in combination with
the masks for recognizing the object. The proposed
use-case innovation is due to 5G infrastructure that
allows transmitting hi-res images to the fleet (moving
and stationary nodes) for coordinated disinfection
response based on distributed object recognition.
A benefit of formation at this stage lies in the
ability of the UGVs to share the observed space by
the onboard cameras in a distributed manner in a
way that will maximize the collectively covered area
(without a given formation this problem of arranging
the onboard cameras of different UVGs to cover a
wide area scene with non-overlapping views would
be highly complex and in some cases unsolvable).
At the same time, recognition algorithms based on
DNNs perform better in the presence of multiple
views of the same object.
The ability to share high-resolution image/videos,
taken by onboard UGV cameras and stationary
nodes, among all the members of the group with the
purpose of performing distributed object recognition
from multiple angles and combining results for the
appropriate disinfection algorithms to be applied,
is what makes this use-case scenario specific to
5G infrastructures. Without 5G high-bandwidth,
low-latency and low-error characteristics shar-
ing this volume of data between the fleet would
be impossible and, in fact, is one of the major lim-
itations in state-of-the-art object-recognition systems.
In the proposed use-case scenario the recognition
is enhanced due to the existence of 5G infrastructure
that permits sharing high-bandwidth, real-time crit-
ical data for distributed object detection. An object
occupying certain bin(s) can be seen from multiple
views performed from all UGV(s) (some with better
view than others). Intermediate recognition steps
depending on the specific algorithm used can also be
transmitted back and forth among the members of
the fleet to better guide the formation for acquiring
more and better images. At the same time, different
UGVs can focus and recognize different objects by
transmitting data and findings to each other, and this
way maximizing the speed of disinfection response if
multiple objects exist around the fleet.
To achieve high-speed response in formation
change for object recognition and subsequent dis-
infection, a progressive algorithm based on shape
analysis from 2D views is proposed here as the first
step in a distributed pipeline of recognition tasks that
will end in DNN recognition from multiple sources
based on boosting techniques. From the 3D bin
mask a UGV can perform a contour extraction from
various views and perform credible shape-based
recognition from a locally kept database of objects.
At the same time, high-resolution images of the
same object taken by onboard fleet UVG cameras
can be used for DNN recognition at each UGV,
with different NN architectures. The combination
of multiple methods in the context of boosting [13],
by means of distributed fleet processing can lead to
enhanced object recognition due to 5G infrastructure.
This, together with the ability to exchange/transmit
high-resolution/bandwidth real-time data back and
forth among the fleet during the recognition and
disinfection tasks, casts the proposed solution as
a significant use-case scenario of 5g networking
technologies.
Regarding the shape-based recognition task, in
particular, it can initially guide the formation. To this
end, the GLS/VAR method [11] is proposed because
its benchmark matching scores are, to the best of
our knowledge, the best one available today for real-
time applications. GLS/VAR is a descriptor of planar
shapes computed on the closed boundary. It uses a
distance map between all pairs of points to represent
shape and the VAR descriptor to define point corre-
spondences between different shapes. The descrip-
tor for each shape is extracted from its boundary to-
gether with its landmark points. The landmark points
are used for cross-shape correspondences at the com-
parison phase. At the time of writing this proposal,
it is, to the best of our knowledge, the fastest method
that can achieve above 70% bulls-eye agnostic score
on the MPEG 1400 benchmark dataset of shapes, thus
it is better suited for large scale search and indexing
applications in the wild. Another important property
of the GLS/VAR descriptor is its global nature and the
resulting resistance to boundary noise[12].
3.1 Proposed Weighted BFT algorithm
A weighted Byzantine Fault Tolerance mechanism is
proposed, considering the hierarchical structure of the
UGV fleet. This paves the way to implement a more
robust decentralized system that would effectively
handle the situation in the case of UGV malfunction
or coordination failures of the fleet that would lead
to non-proper disinfection of an area. This approach
significantly improves the quality of the decision pro-
cess by taking into account inputs coming from vari-
ous types of UGVs in a way that, based on their roles
and importance in their area of responsibility, would
be appropriate. As the influence of such critical nodes
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
243
Volume 6, 2024
will be balanced by the localized insights of individ-
ual UGVs, this weighted approach could lead to a
much more well-rounded decision.
The hierarchical structure that we propose can be
analyzed as follows: Firstly, the leader UGVs (L-
UGVs at the top of the hierarchy, is the first-in-line
coordinator and holds the highest decision-making
weight through strategic overviewing and command
capabilities. Therefore, it is responsible for superor-
dinate mission planning overall resource allocation,
and critical decision-making to ensure that the fleet’s
activities are executed according to the set objec-
tives. Secondly, regional coordinator UGVs (RC-
UGVs) at the mid-level, are intermediate, moderate-
weight UGVs responsible for managing small, spe-
cific subgroups of individual UGVs within defined
regions. They aggregate information from their as-
signed UGVs, process it into more specific forms,
make decisions relevant to the small areas they cover,
and send critical information back to the lead UGV.
At the base of this hierarchy, individual UGVs have
the last word in execution and relate to tasks such as
capturing real-time data, disinfection, and detecting
problems immediately. UGVs bear the least weight
in the decision process but are, nevertheless, front-
line devices for localizing information and grounding
all fleet operations in environmental actualities. In
combination, these three classes of UGVs thus form
a complementary and well-balanced system capable
of strongly decentralized decision-making and robust
execution of assigned missions.
From a mathematical perspective, several key
points should be analyzed to identify the response
as formulated by the majority. Firstly, each UGV
type has a specific weight wi, where i=l, r, u and
wl> wr> wu, for leader UGVs, regional UGVs,
and individual UGVs respectively. The total weight
is calculated by the sum:
W=∑
i=l,r,u
(ni×wi),
where niis the number of specific role UGVs.
To ensure that only a supermajority agreement can
make a decision, the total threshold weight for con-
sensus is set to be equivalent to two-thirds of the total
weight. Promote stability rather than a quick and rash
decision that may be unbalanced, given the state at a
particular moment. This level of support will improve
the decision, as every action taken is most likely well-
thought-out and supported by an array of inputs, while
preventing domination at any level or type of UGV,
leading to balanced participation of different kinds of
vehicles. Therefore the threshold Tis calculated by
T= 2/3 × W, where Wis the total weight.
For a specific action A, the total weight sum of
UGVs supporting such an action within the context
provides the total weighted vote VW(A). Here, the
total weighted vote for a specific action ensures that
influence from each UGV is correctly accounted for
according to its hierarchical weight. The weight in-
dicates the role and importance of UGVs to the fleet.
This uses an indicator function based on UGVs re-
sponses Ri, which is denoted by δ(A, Ri)to sum only
the weights of UGVs that support the given action,
where it equals one if a UGV supports the given ac-
tion and otherwise is equal to 0. This adds weighted
support to any particular stance that underpins the
consequential decision in a fully transparent manner
and balances the strategic oversight, where higher-
weighted UGVs are, with the operational insights,
where lower-weighted UGVs are, to drive rounded
and robust decisions that reflect the overall fleet’s in-
put. The mathematical equation for this is the follow-
ing:
VW(A) = ∑
i=l,r,u
(ni×wi×δ(A, Ri))
In this way, while incorporating the critical real-time
data from the operational UGVs, strategic decisions
get to leverage the much broader perspective provided
by these higher-weighted nodes.
In order to determine the majority response MR,
the algorithm will compare the weighted vote for each
action against the threshold weight. It will iterate
through all possible actions and tests if the weighted
votes of the considered action are meeting or exceed-
ing the threshold weight; once this condition is met,
a choice based on this result will be made as the ma-
jority response to execute. Thus, only those actions
with substantial support from the fleet, as seen by the
weighted consensus, get executed, hence upholding
stability and avoiding the making of unbalanced or
premature decisions:
MR=A,if VW(A)≥ T .
The proposed consensus mechanism is illustrated
in the figure below (Figure 1), while the pseudocode
for this approach is presented by Algorithm 1.
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
244
Volume 6, 2024
Algorithm 1 Weighted Consensus Mechanism for
Emergency Situations in UGV Fleet
1: Initialization:
2: Initialize weights for each type of UGV:
3: Define consensus threshold
4: Initialize the list of UGVs and their roles
5: while True do
6: for each UGV in UGVs do
7: if UGV detects an emergency then
8: Set emergency ←Active
9: Get emergency data from UGV
10: Broadcast the emergency data to all UGVs
11: end if
12: end for
13: if emergency is Active then
14: emergency responses ← ∅
15: for each UGV in UGVs do
16: Each UGV creates a response based on the emergency
17: Add the response to the list of emergency responses
18: end for
19: Calculate the weighted votes based on the emergency responses
20: for each response in emergency responses do
21: UGV _id, action ←response
22: weight ←get_node_weight(U GV _id)
23: if action /∈weighted_votes then
24: weighted_votes[action]←0
25: end if
26: weighted_votes[action]+ = weight
27: end for
28: total_weight ←sum(weighted_votes.values())
29: threshold_weight ←total_weight ×2
3
30: majority_response ←N one
31: for each action in weighted_votes do
32: if weighted_votes[action]≥threshold_weight then
33: majority_response ←action
34: break
35: end if
36: end for
37: if the action in the majority response is ”reroute” then
38: for each UGV in UGVs do
39: Update the route based on the majority response
40: end for
41: else if the action in the majority response is ”assist” then
42: for each UGV in UGVs do
43: if UGV’s location is in the list of assist locations from the ma-
jority response then
44: UGV provides assistance
45: end if
46: end for
47: else if the action in the majority response is ”redistribute” then
48: for each UGV in UGVs do
49: Update task assignment based on the new tasks from the major-
ity response
50: end for
51: end if
52: Set emergency ←F alse
53: end if
54: end while
Figure 1: A flowchart diagram for the proposed BFT
consensus mechanism
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
245
Volume 6, 2024
4 Conclusion
Leveraging its extensive expertise in developing spe-
cialized hardware for unmanned vehicle platforms
(like the fully autonomous UUVs Synoris: [18], [19],
[20]), along with its advanced research in computer
vision, shape analysis, and machine learning, Infili
Technologies S.A. proposes to create a use-case sce-
nario utilizing cutting-edge 5G-enabled recognition
and formation technology. The new system can per-
form recognition and subsequent disinfection tasks in
hospitals from mainly visual sensor data by appropri-
ate and controlled formation positioning of a fleet of
UGVs by using their onboard adjustable UV lamps,
in a distributed and optimal manner, empowered by
5G connectivity between the system’s nodes.
Acknowledgment:
It is an optional section where the authors may write
a short text on what should be acknowledged
regarding their manuscript.
References:
[1] Joseph Redmon and Ali Farhadi, YOLOv3: An
Incremental Improvement, arXiv, 2018
[2] O. Ronneberger, P. Fischer, and T. Brox. U-net:
Convolutional networks for biomedical image
segmentation. In International Conference on
Medical Image Computing and
Computer-Assisted Intervention (MICCAI),
pages 234–241. Springer, 2015.
[3] S. Yu, B. Park, and J. Jeong. Deep iterative
down-up CNN for image denoising. In CVPR
Workshops, 2019.
[4] B. Park, S. Yu, and J. Jeong. Densely connected
hierarchical network for image denoising. In
CVPR Workshops, 2019.
[5] [15] G. Huang, Z. Liu, L. Van Der Maaten, and
K. Q. Weinberger. Densely connected
convolutional networks. In CVPR, pages
4700–4708, 2017.
[6] [14] K. He, X. Zhang, S. Ren, and J. Sun. Deep
residual learning for image recognition. In
CVPR, pages 770–778, 2016.
[7] [43] Y. Zhang, Y. Tian, Y. Kong, B. Zhong, and
Y. Fu. Residual dense network for image
super-resolution. In CVPR, pages 2472–2481,
2018.
[8] Y. Zhang, Y. Tian, Y. Kong, B. Zhong, and Y. Fu.
Residual dense network for image restoration.
arXiv preprint arXiv:1812.10477, 2019.
[9] Y. Zhang, K. Li, K. Li, L. Wang, B. Zhong, and
Y. Fu. Image super-resolution using very deep
residual channel attention networks. In ECCV,
pages 286–301, 2018.
[10] M. Haris, G. Shakhnarovich, and N. Ukita.
Deep backprojection networks for
super-resolution. In CVPR, June 2018.
[11] Konstantinos A. Raftopoulos and Stefanos D.
Kollias. The Global–Local transformation for
noise resistant shape representation. Computer
Vision and Image Understanding,volume = 115,
number = 8, pages = 1170-1186, year = 2011
[12] Konstantinos A. Raftopoulos, Stefanos D.
Kollias, Dionysios D. Sourlas, and Marin
Ferecatu. 2018. On the Beneficial Effect of
Noise in Vertex Localization. Int. J. Comput.
Vision 126, 1 (January 2018), 111–139.
https://doi.org/10.1007/s11263-017-1034-6
[13] Chen, Tianqi and Guestrin, Carlos. XGBoost:
A Scalable Tree Boosting System. Proceedings
of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data
Mining, pages=785–794, year=2016,
organization=ACM,
doi=10.1145/2939672.2939785
[14] A. Artusi and K. A. Raftopoulos, ”A
Framework for Objective Evaluation of Single
Image De-Hazing Techniques,” in IEEE Access,
vol. 9, pp. 76564-76575, 2021, doi:
10.1109/ACCESS.2021.3082207.
[15] Chen, Chang Wen et al. ”Deep Boosting for
Image Denoising.” European Conference on
Computer Vision (2018).
[16] Xu, Jun et al. ”External Prior Guided Internal
Prior Learning for Real-World Noisy Image
Denoising.” IEEE Transactions on Image
Processing 27 (2017): 2996-3010.
[17] Zhang, K., Zuo, W., Zhang, L. (2018).FDNet:
Toward a fast and flexible solution for
CNN-based image denoising. IEEE Transactions
on Image Processing, 27(9), 4608–4622. DOI:
[10.1109/TIP.2018.2839891]
(https://doi.org/10.1109/TIP.2018.2839891)
[18] Nikolaos Papadakis, Dimitris Paraschos,
”Synoris: an Unmanned Underwater Platform
Based on Hydrophone Arrays for Detection and
Tracking From Sound Signatures,” International
Journal of Circuits, Systems and Signal
Processing, vol. 16, pp. 780-786, 2022, DOI:
10.46300/9106.2022.16.96.
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
246
Volume 6, 2024
[19] Dimitrios Paraschos, Nikolaos Papadakis,
Evangelos Mantas, 12th NMIOTC Annual
Conference 1-2 June 2021, SOUDA BAY
GREECE, SYNORIS – The ”Chariot” of
Modern Maritime Warfare, NATO Matitime
Interdiction Operations journal Issue 22 (1st
Issue 2021) ISSN 2242-441X.
[20] Paraschos D, Papadakis N. K. Autonomous
underwater vehicle challenge: design and
construction of a medium-sized, AI-enabled
low-cost prototype. The Journal of Defense
Modeling and Simulation. 2024;21(3):269-281.
doi:10.1177/15485129211027236.
[21] Castro, Miguel, and Barbara Liskov. ”Practical
byzantine fault tolerance.” OsDI. Vol. 99. No.
1999. 1999.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present re-
search, at all stages from the formulation of the prob-
lem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflicts of Interest
The authors have no conflicts of interest to
declare that are relevant to the content of this
article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International , CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
International Journal of Electrical Engineering and Computer Science
DOI: 10.37394/232027.2024.6.28
Nikolaos K. Papadakis
E-ISSN: 2769-2507
247
Volume 6, 2024
