Combining Clustering Algorithms to Extract Symmetric Clusters from

Noisy Data, Applied to Parking Lots

Gyulai-Nagy Zolt

an-Valentin

Department of Computer Science, Babes¸-Bolyai University,

1, M. Kogalniceanu Street, 400084, Cluj-Napoca, Romania

Keywords:

Clustering, SOM Clustering, HDBSCAN, K-Means, Real-Time Application, Automation, Parking Lot Space

Detection, Feature Extraction.

Abstract:

The paper presents an approach for detecting symmetrical clusters in noisy data, using parking space de-

tections as a real-world example. The paper proposes a plug-and-play solution that uses camera systems to

automatically detect parking spaces and provide metrics about availability and accuracy. The approach uses

clustering algorithms and image detection for data acquisition and mapping, which can be easily adapted to

any application that requires geometrical data extraction. The paper also presents the different phases involved

in mapping parking spaces and the challenges that need to be addressed. Overall, the proposed approach can

beneﬁt both parking lot administrators and drivers by providing real-time information on available parking

spaces and reducing emissions, fuel costs, trafﬁc, and time spent searching for a spot.

1 INTRODUCTION

Detecting where there are free spaces in a parking

lot has always been a complex problem. In previ-

ous years, as cars play a great role in our day-to-day

means of transportation, cities got agglomerated and

therefore, it is harder and harder to ﬁnd free spaces

in parking lots. One might spend more time search-

ing for an available space in big cities than traveling

to the destination. We need to also take into account

the pricing of leaving your car at a lot. When in high

demand due to having a preferential placement in the

central area, the price might be a bit higher when com-

pared to other options. One might park farther from

their destination to avoid paying higher fees. Lately,

there have been a few initiatives to incentivize peo-

ple to use public transportation for daily activities, but

this does not seem to solve the problem of parking

lot availability, as space is limited in crowded cities.

(Cooke and Behrens, 2017) It would be helpful to just

open an application and see in real-time where we

might have the chance to ﬁnd available spaces. This

would also contribute to reduced emissions, therefore

a greener environment, cost savings on fuel, reduced

trafﬁc, and time efﬁciency, due to not being required

to search for a space.

https://orcid.org/0000-0001-6506-4679

This paper will focus on combining clustering al-

gorithms to detect parking spaces based on positional

data. Unlike other approaches, we provide a plug-

and-play solution that does not require user-provided

information at all, provides metrics and insights, re-

lies on potentially existing equipment, can automati-

cally adapt to any topology, and does not require com-

plex setup or hardware.

2 LITERATURE REVIEW

In the past, the topic of detecting the occupancy of

parking lots has been answered with different kinds

of approaches. One of the most straightforward ones

is manually marking the spaces and using simple im-

age recognition on cars to predict occupancy. This has

evolved into a race where applications try to detect the

spaces automatically using the available data from im-

ages. This approach uses features that might delimit

a space, like markers placed on the ground (Acharya

et al., 2018). Other similar approaches try to optimize

the images further to reduce the impact of lightning

conditions or glare (Huang et al., 2022). Further re-

search has been conducted to rely only on the position

of the vehicles. Although a camera is still required,

using positional data also has advantages when con-

sidering unmarked spots. This involves using cluster-

362

Zoltán-Valentin, G.

Combining Clustering Algorithms to Extract Symmetric Clusters from Noisy Data, Applied to Parking Lots.

DOI: 10.5220/0012624400003687

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

In Proceedings of the 19th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2024), pages 362-369

ISBN: 978-989-758-696-5; ISSN: 2184-4895

ing and predeﬁned information about the lot (Grbi

and Koch, 2023). Another more popular approach,

due to its applicability to a large variety of topolo-

gies, is to use sensors placed in the ground or above

the cars to detect occupancy. This approach has the

advantage of not requiring a camera. Therefore, in-

formation that might be private to drivers but embed-

ded in the images, like car plates, doesn’t need to be

trimmed off, and potential security breaches are ex-

cluded. It requires more manual setup, but once done,

it provides accurate data (Zhang et al., 2015). Mod-

ern vehicles are being equipped with multiple types

of sensors and cameras. Some approaches try to use

them to provide information about parking lots. This

has the advantage of not requiring the installation of

new sensors (Luo et al., 2017).

Previous research has explored the combination

of clustering methods, revealing that leveraging re-

sults from multiple algorithms can indeed enhance

accuracy in distinguishing members across multiple

clusters. (Khedairia and Khadir, 2022) Another pa-

per delves into the effectiveness of amalgamating K-

means with Genetic Algorithms, leading to processes

that are both more accurate and efﬁcient. (Zeebaree

et al., 2017) Additionally, research has demonstrated

that combining weak clustering algorithms can fur-

ther enhance the accuracy of the outcomes. (Topchy

et al., 2003). In conclusion we can state that there

can be beneﬁts in terms of performance and accuracy

when well known algorithms are combined.

Furthermore, a noteworthy aspect pertinent to our

methodology is the intriguing research where it was

proven that by inputting initial recognition data from

image processing into a neural network with the inten-

tion of retraining it, we can attain improved classiﬁ-

cation performance (Doulamis et al., 2000). All these

ﬁndings contribute to our methodology and provide

the basis for our contribution.

3 PROBLEM STATEMENT AND

RESEARCH QUESTIONS

Mapping of parking spaces is a complex operation

and requires multiple phases to achieve adequate re-

sults without relying on any kind of user input. The

adaptability to changes in the topology and providing

metrics also means that we need to do continuous pro-

cessing, and the results need to be as accurate as they

can get between each mapping phase to provide value

in real-world applications. In the following subsec-

tions, we are going to provide details on each phase

and the problem that they need to solve.

3.1 Data Acquisition and Preprocessing

First of all, we need to acquire the data that will be

used further in the next phases. In our application, we

want to rely on camera images as lots of parking lots

already have them placed due to security measures.

This also means that images should cover most of the

available area. In this paper, we will exclusively fo-

cus on scenarios where a single camera constructs the

topology of a parking lot. This choice doesn’t com-

promise the plug-and-play functionality of the sys-

tem; rather, it maintains it for single-camera setups.

However, future enhancements can extend this capa-

bility to derive topology from multiple camera angles,

resulting in more precise data.

Most cameras are placed higher above ground to

cover as much space as they can. The angle of view

and lens distortion are factors that we need to take into

consideration. We can work with non-uniform data

as it is going to be presented in the next sections, but

transforming the data is going to provide further bene-

ﬁts in distinguishing symmetrical features. Therefore

the ﬁrst research question is phrased as follows:

RQ1 What is the most appropriate transformation for

the data to illustrate the symmetrical features?

3.2 Data Denoising

By using the detected cars in the image, we are go-

ing to have lots of noise. This is because even if we

process the images from a feed at a given time inter-

val, there is going to be movement of departing and

arriving cars.

We also need to take into consideration that we

might have such areas in a parking lot where illegal

parking occurs. By illegal parking, we don’t mean it is

punishable by law to leave the car there; that’s a factor

that might be unreliable to be extracted purely from

images, and it’s not in our scope. We actually mean

that space is not designated to be used as a parking

space, and even so, people leave their cars there as it

won’t impact other drivers’ access to other spaces.

This would provide valuable insight for drivers to

determine by their arrival how likely it is to ﬁnd their

chosen space occupied. Given the previous informa-

tion, we can conclude our interests:

RQ2. What algorithm can denoise and regularize the

data?

3.3 Determining the Number of Parking

Spaces

We need to take into consideration lots of variables in

the detected data when we want to determine the num-

Combining Clustering Algorithms to Extract Symmetric Clusters from Noisy Data, Applied to Parking Lots

363

ber of available parking spaces. Depending on how

people park their cars, in the more accessible spaces,

we are going to have much denser data, whereas in ar-

eas that are like a fallback to if the main places are oc-

cupied, we are going to have much fewer data points.

Therefore if a parking lot is frequently at its peak ca-

pacity, we have a greater uniformity in data, but we

can’t solely rely on that as we want our approach to

work on less frequented lots too. If the detected data

is not dense enough, we might want to consider them

as noise, but in time as the samples increase, it is de-

sired that they get to be detected as valid points.

A small memory-like approximation is also de-

sired. As the topology of a parking space does not

really change too often and the camera is also kept in

the same place, remembering on a short term where

there were the previous detections it’s a helpful fea-

ture. We can also extract further data from these de-

tections and we can make sure that if, for instance, the

camera deteriorates and it’s noticed in a short period

of time, we still have stable estimations based on past

data.

RQ3. What algorithm can delimit clusters via data

frequency and can keep them in memory?

3.4 Parking Space Assignments and

Real-Time Detection

As a ﬁnal phase, we want to detect a lot’s occupancy

in real time. For this, we need an estimate for the size

of each parking space and their center point. We also

need to detect the center points of the clusters in such

a way that they’re placed in the most dense area. It

is important to delimit the parking spaces by priori-

tizing the Euclidean distances between them to avoid

putting more clusters than needed over areas that are

highly dense in data.

RQ4. What algorithm is capable of accurately posi-

tioning centroids based on Euclidean distance?

Figure 1: Steps of data processing.

4 METHODOLOGY

We will use a multi-step process to provide real-time

data about the occupancy of parking spaces. We can

differentiate two streams in this process: one is re-

sponsible for creating the parking lot model, and the

second is providing real-time metrics and data. In the

following subsections, we are going to describe them.

An overview of the following steps can be found in

Figure 1.

4.1 Data Acquisition and Preprocessing

For this section, we are going to use the PKLot dataset

(de Almeida et al., 2015). It provides images of a

parking lot over a longer period of time, and it covers

multiple scenarios like daytime images, nighttime im-

ages, and different types of weather conditions. The

angle of the camera is positioned in such a way that

it covers the lot. It contains scenarios where the cars

are not only in preferred positions, like stationed in a

parking lot, but also when they are moving in and out

of the parking spaces. It also contains images where

people leave cars in unmarked spots. Basically, it con-

tains most scenarios that are going to be covered when

we use a camera stream and pick a few images at pre-

deﬁned time distances to be processed. As we want to

cover real-world scenarios, this is a viable database.

To detect the cars in the images, we are going to

use YOLOv8 (8th version of YOLO model). In the

past few years, it has been one of the best solutions

when talking about real-time detection of different

types of objects (Terven and C

ordova-Esparza, 2023).

The accuracy of detections is good, the lighting con-

ditions do not have such a signiﬁcant impact on the

results, and it has conﬁgurable models depending on

what accuracy we want to achieve. It is a good can-

didate to be used on small devices like Raspberry Pi’s

that can be extended with camera modules. This is

important because, as previously mentioned, data pri-

vacy is an essential factor when images are involved.

Using a device that can do the processing on the spot

reduces the risk of leaking additional, potentially pri-

vate, information via internet connections. We as-

sume that having the ﬂexibility to choose between dif-

ferent types of models can bring further advantages

in the cost optimization domain. Going forward we

are going to use the YOLOv8n (8th version of YOLO

model with 3.2 million parameters) model, but we are

going to include in the tests other models too.

Once the cars are detected in the images, we are

going to store the center points of the bounding boxes.

These points need to be processed further to better

indicate the symmetry between the parking spaces.

Due to a lack of raw dept information, the eagle view

seems to be the desired way of viewing the data, as the

dimensionality of the points is reduced from 3D to 2D

by trimming out the depth of ﬁeld. We can achieve

this by various methods, but via recent research, we

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

364

can easily do so with geometric approaches (Abbas

and Zisserman, 2020). We can use the vertical van-

ishing point and ground plane vanishing lines to com-

pute a homography matrix that will transform the

data. These can be computed by using a Convolu-

tional Neural Network (CNN) trained in this scope.

The network is trained on a large amount of synthetic

images with known camera conﬁgurations.

After the perspective transformation is applied, we

are going to pass the data to the next phase.

In the next step, we are going to use HDBSCAN

clustering. It extends DBSCAN into a hierarchical

clustering algorithm that is able to ﬁnd clusters with

varying densities. The only parameter that we are go-

ing to use as input is the number of minimum cluster

sizes. As we don’t have an exact number for that,

we are going to iterate over a few values to remove

the noise. If our data is entirely evenly distributed

in the parking lot, we could use 2882 as the starting

value, and the resulting cluster number should be 24.

This won’t be the case as the spaces on the right side

of the image are less frequented, and the YOLOv8n

model will have fewer detections due to image dis-

tortion. We are going to use this value as a starting

point, and we are going to decrease it until we get to

at least 24 detected clusters. Once we have that value,

we consider a simple rule to eliminate clusters that

are too large and present irregularities. For each itera-

tion going forward, we calculate the average width of

the clusters. Everything that is larger than the average

times two is going to be ﬂagged. We chose this met-

ric because we assume that the spaces are symmetric,

meaning that if the width of a cluster is double com-

pared to the average, it can be split into two additional

clusters. We decrease the minimum cluster size by

100 and clusterize the data again. The 100 is a num-

ber that provides enough granularity in the iterations.

A smaller number can be used, but it would result in

more iterations. In case a larger cluster breaks into

smaller ones, and the total number of ﬂagged clus-

ters also decreases, having noise points increasing,

we remove the previously ﬂagged clusters. We aim

to ﬁnd scenarios where clusters are not separated and

symmetrical enough, and they are connected in some

grade. This ensures we remove clusters that are too

large and can’t be clearly classiﬁed as parking spaces.

To better understand the iterations, the following bul-

let points summarize the steps: (1) Determine the av-

erage max width of clusters. (2) Flag the clusters that

are larger than the double of average width. (3) If

there are fewer ﬂagged clusters compared to the pre-

vious step and there is additional noise, remove the

previously ﬂagged clusters. (4) Decrease min cluster

size with 100 (or any smaller number).

The process stops once we get to a minimum clus-

ter size smaller or equal to the decrease value (100) or

when the number of detected clusters on the original

data is closest to the maximum number of detected

cars in an image, which is 64 in this case.

The presented method deﬁnes constraints for up-

per and lower boundaries for the denoising algorithm,

efﬁciently removing noise and clusters with ambigu-

ous symmetry. As we can see in the images, the num-

ber of resulting clusters is approximately close. How-

ever, further processing needs to be done to deﬁne the

exact number of parking spaces due to the wide vari-

ety of densities.

4.2 Determining the Number of Parking

Spaces

This phase aims to determine the number of parking

spaces that are available in the parking lot. For this

purpose, we are going to use the self-organizing maps

(SOM) (Miljkovi

c, 2017).

Due to other neural network properties, we will

have the beneﬁt of short-term memory. This is useful

if the network is retrained with faulty data; previous

information will leave its mark on the model, and we

might be able to detect previously presented spaces

that have been removed from the lot for a short period

of time. This provides further metrics for the input

data in determining the probability of detecting park-

ing spaces.

Once we have trained the SOM model, we are go-

ing to extract the distance map. This enables us to

visualize the distance of nodes compared to neighbor-

ing ones. We are going to map the values to white and

black, white representing the most distant nodes and

black the closest ones. To further extract information

from the image, we are going to apply a few image

processing algorithms to make everything more inter-

pretable. We apply an adaptive histogram equaliza-

tion to increase the contrast. This is also useful be-

cause where we have color information, they are go-

ing to be more accentuated, but where we have black

shades, the new values are still going to be close to

each other. (Pizer et al., 1987) This approach has a

slight drawback; even though the information is eas-

ier to interpret, the noise can get overampliﬁed. To

solve this issue, we are going to apply a local mean

ﬁlter with a low neighboring pixel check to the image.

This ensures that noise is reduced, and we can easily

extract the number of clusters shown in the image.

To do so, we will use a simple watershed algorithm

where the number of clusters will represent the num-

ber of parking spaces in our lot. The algorithm places

the starting points for the ﬁll process in the areas of

Combining Clustering Algorithms to Extract Symmetric Clusters from Noisy Data, Applied to Parking Lots

365

the image that have the least gradient change.

Now that we have the actual number of spaces and

we have access to the denoised data, we can continue

to the following phase.

4.3 Parking Space Assignments

At this step and also the ﬁnal processing step, we

have access to the denoised data and the number of

available spaces. We need to deﬁne the cluster cen-

ters precisely and transform them back to the original

information presented in the image to be visually in-

spected. For this purpose, we are going to use the

KMeans algorithm. It is highly efﬁcient in assigning

each data point to the nearest cluster and minimizing

distance variances in clusters based on the Euclidean

distance. Even though it is a computationally de-

manding process, the partitioning converges quickly

to the local minimum (Ahmed et al., 2020). Only one

input parameter is required, that being the number of

clusters, but we already know that value from previ-

ous steps. With this algorithm, we have the assurance

that the centroids are placed in the most optimum po-

sition.

Once we have all the data, we store them for

real-time detection and retraining in case it’s needed.

We store the transformed data (with nodes marked as

noise, parking slot, and illegal parking), the median

width of a cluster, the SOM model, including the dis-

tance map, the number of detected clusters, and the

transformed centroids, including the real-world cen-

troids.

4.4 Finetuning Our Model

Once we have all the clusters placed, it is essential

that we train our YOLOv8 model to increase the ef-

fectiveness of our real-time processing. Our initial de-

tections are going to have a below-average accuracy.

To improve this, we are going to take the bounding

boxes of the detected cars that overlap with the previ-

ous clusters, and we are going to retrain the YOLOv8

model based on them. We will reuse the already exist-

ing label of a car and assign it to the previous bound-

ing boxes then feed it to the YOLOv8 training pro-

cess.

This ensures that we improve the accuracy of our

detections and that we customize the model to ﬁt our

parking lot, by retraining the YOLOv8 model once.

Previous detections are going to be stored. This ap-

proach is suitable to increase the accuracy only if

there is enough data to be processed, but in case there

are not enough detections, the system will wait until

enough data is collected.

4.5 Real-Time Detection

To detect whether a parking space is occupied and

provide metrics about the lot, we will use a short pro-

cessing of live feeds to provide real-time data. The

process starts with selecting images from the feed at

a predeﬁned time distance. Depending on how accu-

rately we want to access information, it can be deﬁned

as a shorter or longer period. We will propose time-

frames of one minute as it will be accurate enough

for scenarios where drivers are in search of spaces.

On the selected images, we are going to apply the

YOLOv8 classiﬁcation with any kind of model de-

pending on the topology of the parking lot. If there

are only a few cars and they are all clearly visible,

YOLOfv8n would be enough. Otherwise, it makes

more sense to use a more performant model. The

bounding box centroids from the detected vehicles are

then transformed into the eagle view. These are going

to be stored in a backlog for further retraining and

passed to the next step. Using the centroid positions

of the parking spaces and the bounding box centers,

we can easily conclude which one is the closest based

on the Euclidean distance.

We are going to extract further metrics from the

data to evaluate the accuracy of the previous classiﬁ-

cation. The ﬁrst metric is the overlap with other park-

ing spaces. In case we detect a car centroid in the

range of multiple parking spaces, where double the

average cluster width is still in range for all of them,

we can conclude that each space can be wrongly occu-

pied. This is an irregularity that might still allow an-

other car to park near the wrongly parked car. There-

fore, we are going to treat it as a ﬂag for each space.

The next metric is given by the removed clusters that

were not uniform. If a car is detected multiple times

in the same spot, we can say it is stationary. If we

extend this constraint so that the centroid of the car

overlaps with the area of a removed cluster, we can

extract a ﬂag that indicates a car parked in an illegal

spot. The total number of such ﬂags helps to detect

the overcrowdedness of a parking lot.

The last metric is the conﬁdence factor for all pre-

vious classiﬁcations and ﬂags. By using the SOM dis-

tance map translated into the pixels of the image, we

can generate a conﬁdence factor based on the distance

of the node that is being triggered inside the model.

The red boxes mark positions where the ﬁltered

parking spaces have been detected, and the dots rep-

resent by color the mapped distance of the SOM node.

If it is red, it means it is a very distant node, whereas

if it’s blue, it means it’s from a crowded area. We

will use another variable, whether the node is a win-

ner node or not. We say that a node is a winner if it is

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

366

Table 1: Accuracy of detected spaces.

Subset Ofﬁcial spaces Detected Spaces Actual spaces Accuracy

UFPR04 28 30 26 95.56%

UFPR05 45 45 43 92.86%

PUCPR 100 excluded excluded excluded

triggered for the training data.

For the parking spot classiﬁcation values, we can

state that in case a spot is occupied, and the triggered

node is a winner node with a low distance value, we

are conﬁdent that the detection is accurate; otherwise,

it is not.

In case a parking space or more of them are

wrongly occupied, and the triggered node is either not

a winner node or the distance value of it is higher than

0.5, we are conﬁdent that the detection is accurate;

otherwise, it is not.

In the last scenario where a car is marked as ille-

gally parked, we can state that we are conﬁdent about

the detection in case the activated node has a lower

distance value than 0.5. Unfortunately, due to the fact

that SOM requires higher and lower boundaries for

the input data to be segmented, we can not conclude

that we have no conﬁdence in our detection. There-

fore, we are only left with a positive conﬁdence value

if applicable.

Lastly, we are going to deﬁne the cases where the

system detects a change in the topology and will re-

iterate the training steps to gain improved accuracy.

If a car can not be classiﬁed in any of the previously

mentioned categories and is detected to be stationary

for a longer time, on multiple occasions, the system

will grab the historical data from the last training pro-

cess and retrain itself. Another case is going to trigger

the same behavior, and that is when, on multiple oc-

casions in the same spots, cars are detected to occupy

more than one space by being wrongly parked.

Whenever such a scenario occurs, until the prede-

ﬁned amount of repetitions is not reached, the system

is going to show a ﬂag that states that lot maintenance

or topology change has been detected. The retraining

process ensures future prooﬁng and automated main-

tenance for the system.

To achieve optimal results for the system, it is nec-

essary to accumulate data over a period of at least one

month. This ensures that an adequate number of sam-

ples are collected from both high and low trafﬁc times

and the YOLOv8 model can be retrained. Results

with sub optimal precision can be observed within the

ﬁrst week from deployment.

5 RESULTS

Our results, by using the previously mentioned ap-

proach, can make use of a short interpretation before

we dive into the testing values and comparison. As il-

lustrated in the previous images, the system is able to

automatically recognize parking spaces and detect the

occupancy of spaces in real-time while also providing

additional metrics. Interestingly, even though we de-

tected 45 parking spaces, which is the exact number

of spaces seen on the camera, we missed two spots but

added two unmarked spots. The two spaces that have

been left out are at the top of the image. The ﬁrst one

has been removed due to the constraints of the HDB-

SCAN algorithm. This means the distance and den-

sity of that cluster regarding the symmetry between

clusters were not met. The SOM algorithm removes

the second space. This is due to the fact that we do

not have enough samples in that cluster to be classi-

ﬁed as a completely different cluster. This can be be-

cause we used the YOLOv8n model, which adds fur-

ther noise to the data by having less accurate bound-

ing box placements and fewer detected cars on the

images. But this also demonstrates that the process

works as designed and has a high ﬂexibility.

The two spaces that were added satisfy the con-

straints for each step. Now the question would be:

would it be accurate to classify those two spaces as

an error if they look and act as regular parking spaces

and no regulation is applied to use them? These detec-

tions could also provide insights into further parking

topologies that can be implemented to enable more

usable spaces; therefore, for now, we are not going to

treat it as an error but as an insight.

5.1 Testing

In the methodology section, we saw the capabilities

of the approach and the beneﬁts. To further com-

pare the approach with other implementations and

validate the beneﬁts, we will ﬁrst compare the best

and worst YOLOv8 models available used in our pro-

cess. From the parking spaces point of view, although

YOLOv8n was able to detect exactly 45 spaces, two

of them were unmarked spaces. This results in an

accuracy of 95.(5)% for space detection in case we

Combining Clustering Algorithms to Extract Symmetric Clusters from Noisy Data, Applied to Parking Lots

367

count only the 43 actual spots. In our tests for detect-

ing the number of parking spaces, there were no dif-

ferences between YOLOv8n and YOLOv8x 8th ver-

sion of YOLO model with 68.2 million parameters),

although, in the denoising process, there were differ-

ences at which step the clusters got removed. This

empowers the ﬂexibility of the system by providing

support and accuracy for heavier and lighter detec-

tion models. In 1 we can see the accuracy of de-

tecting parking spaces. The detected spaces column

gives us the total detected spaces, whereas the actual

spaces column is calculated by subtracting the num-

ber of parking spaces that are found in addition to the

actual number of spaces. For example, if we detect

45 spaces and two additional spaces are found that

are not included in the ofﬁcial spaces, we are going

to use only 43. The accuracy is going to be given by

the percentage between the ofﬁcial spaces and actual

spaces columns. The rows represent different subsets

of the PKLot database, where UFPR04 and UFPR05

are different camera views of the same parking lot,

and PUCPR is a different lot.

We need to mention that for UFPR04, the sys-

tem actually detects 30, of which 4 are actual parking

spaces, but trees obscure them; therefore, the ofﬁcial

database did not include them in the visible parking

spaces count. 2 spaces are not found at the edge of

the image due to a lower amount of samples. In this

subset there is a camera movement that results in the

shift of the perspective. As a unique feature, our sys-

tem automatically detects and adapts this change. For

a short period of time, it will retrain itself, and after

a few steps where it acquires samples, it will give the

same results as before the perspective shift.

The PUCPR subset is excluded from the test due

to the unreliable amount of detections provided by

YOLOv8. The images can be further segmented, or

YOLOv8 can be retrained (Carrasco et al., 2021) to

provide more accurate detections, but this requires

further development to be applicable for a general-

purpose application, and we chose not to include this

scenario as we want to focus in this paper on the clus-

tering approaches and methodology.

To further provide insights we compare our ap-

proach with two other classiﬁers. For this, we are go-

ing to use the real-time detection feature of our appli-

cation and compute the accuracy values by iterating

each image as if it were a video stream. We consider

it an accurate detection if a car is parked in a parking

space detected by our system, and that space is also

included in the ofﬁcial spaces. In Table 2 we can see

the comparison with mAlexNet (Amato et al., 2017)

and CarNet (Nurullayev and Lee, 2019).

Out of the box, the YOLOv8x model provided a

Table 2: Real-time detection accuracy.

Subset mAlexNet CarNet Ours

UFPR04 99.54% 95.60% 84.97 %

UFPR05 99.49% 97.60% 86.08 %

PUCPR 99.90% 98.80% excluded

72% accuracy when compared to the ground truth.

The impact of the improved model is consistent, as

seen in the previous table, and the test results were

measured after training the YOLOv8n model only for

two epochs with a precision coefﬁcient above 0.7 for

detections. This demonstrates the ﬂexibility of this

approach and its applicability to devices that do not

have too much processing power. As seen in (Car-

rasco et al., 2021), the accuracy after retraining the

YOLOv8 model can get up to 96.34%.

6 CONCLUSION AND FUTURE

WORK

In the current research, we have developed a reﬁned

data processing framework speciﬁcally tailored to ad-

dress the challenges associated with geospatial data

acquired from camera systems. This new framework

incorporates denoising algorithms that improve the

quality of the collected data, can extract metrics, and

features a short-term memory, facilitating the effec-

tive clustering of parking lot data based on transient

usage patterns.

Furthermore, we have explored the capacity for

performance enhancement in neural network mod-

els through a process of retraining. We observed a

substantial increase in model accuracy and reliability

post-retraining by initiating the process with a base-

line neural network model conditioned on the ini-

tial dataset (comprising visual recognition of park-

ing space occupancy). This approach emphasizes the

practical beneﬁts of our processing framework in op-

timizing neural network applications for categorizing

and predicting parking lot usage.

We can conclude that the presented approach is a

good ﬁt for general-purpose applications. It is highly

ﬂexible in terms of platforms where it can be hosted,

and it’s also adaptable to the topology of the parking

lot while providing further metrics. It can be applied

to also other types of purposes like pedestrian trafﬁc

analysis or, with the change of the input stream even,

seismical data analysis, as its key feature is to detect

the symmetry between data clusters.

Despite appearing less effective than other ap-

proaches, the method proposed in this article priori-

tizes ﬂexibility and user-friendliness. Its accuracy is

expected to improve over time. One of its main ad-

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

368

vantages lies in its automatic adaptability and the ad-

ditional metrics it offers for assessing various condi-

tions within a parking lot. While further research may

lead to enhanced accuracy, the current methodology

can represent added value when applied to business

scenarios.

Based on the accuracy of up to 95.56% of our ap-

proach, we can answer the research questions in the

following ways:

• RQ1: the eagle view is the most appropriate

transformation of the data to understand the sym-

metry between topological data better.

• RQ2: HDBSCAN combined with a descending

min cluster sample value can denoise data and re-

move irregularities.

• RQ3: SOM clustering can better delimit clusters

based on data frequency and keep track of previ-

ous clusters for a short time.

• RQ4: K-means algorithm can place the cluster

centers accurately based on Euclidean distance.

One of the improvements for future work might

be the approximation of delimiting spacing lines be-

tween the parking spaces. As long as all spaces are

marked, placing a separator line between them should

be possible. This would provide further possibili-

ties for evaluating the existence of additional parking

spaces that can exist at the edge of the image.

Another topic for future work would be to evalu-

ate the accuracy of the clustering approach after the

YOLOv8 model has been trained. It should be able

to provide further insights into the topology of data

points.

REFERENCES

Abbas, A. and Zisserman, A. (2020). A Geometric Ap-

proach to Obtain a Bird’s Eye View from an Image.

Acharya, D., Yan, W., and Khoshelham, K. (2018). Real-

time image-based parking occupancy detection using

deep learning.

Ahmed, M., Seraj, R., and Islam, S. M. S. (2020). The k-

means Algorithm: A Comprehensive Survey and Per-

formance Evaluation. Electronics, 9(8).

Amato, G., Carrara, F., Falchi, F., Gennaro, C., Meghini, C.,

and Vairo, C. (2017). Deep learning for decentralized

parking lot occupancy detection. Expert Systems with

Applications, 72:327–334.

Carrasco, D., Rashwan, H., Garc

ıa, M., and Puig, D. (2021).

T-YOLO: Tiny vehicle detection based on YOLO and

multi-scale convolutional neural networks. IEEE Ac-

cess, PP:1–1.

Cooke, S. and Behrens, R. (2017). Correlation or cause?

the limitations of population density as an indica-

tor for public transport viability in the context of a

rapidly growing developing city. Transportation Re-

search Procedia, 25:3003–3016. World Conference

on Transport Research - WCTR 2016 Shanghai. 10-

15 July 2016.

de Almeida, P. R., Oliveira, L. S., Britto, A. S., Silva, E. J.,

and Koerich, A. L. (2015). PKLot – A robust dataset

for parking lot classiﬁcation. Expert Systems with Ap-

plications, 42(11):4937–4949.

Doulamis, A., Doulamis, N., and Kollias, S. (2000). On-line

retrainable neural networks: improving the perfor-

mance of neural networks in image analysis problems.

IEEE Transactions on Neural Networks, 11(1):137–

155.

Grbi

c, R. and Koch, B. (2023). Automatic vision-based

parking slot detection and occupancy classiﬁcation.

Expert Systems with Applications, 225:120147.

Huang, C., Yang, S., Luo, Y., Wang, Y., and Liu, Z. (2022).

Visual Detection and Image Processing of Parking

Space Based on Deep Learning. Sensors, 22(17).

Khedairia, S. and Khadir, M. T. (2022). A multiple clus-

tering combination approach based on iterative voting

process. Journal of King Saud University - Computer

and Information Sciences, 34(1):1370–1380.

Luo, Q., Saigal, R., Hampshire, R., and Wu, X. (2017). A

Statistical Method for Parking Spaces Occupancy De-

tection via Automotive Radars. In 2017 IEEE 85th Ve-

hicular Technology Conference (VTC Spring), pages

1–5.

Miljkovi

c, D. (2017). Brief review of self-organizing maps.

In 2017 40th International Convention on Information

and Communication Technology, Electronics and Mi-

croelectronics (MIPRO), pages 1061–1066.

Nurullayev, S. and Lee, S.-W. (2019). Generalized Parking

Occupancy Analysis Based on Dilated Convolutional

Neural Network. Sensors, 19(2).

Pizer, S. M., Amburn, E. P., Austin, J. D., Cromartie, R.,

Geselowitz, A., Greer, T., ter Haar Romeny, B., Zim-

merman, J. B., and Zuiderveld, K. (1987). Adaptive

histogram equalization and its variations. Computer

Vision, Graphics, and Image Processing, 39(3):355–

368.

Terven, J. R. and C

ordova-Esparza, D.-M. (2023). A Com-

prehensive Review of YOLO: From YOLOv1 and Be-

yond.

Topchy, A., Jain, A., and Punch, W. (2003). Combining

multiple weak clusterings. pages 331– 338.

Zeebaree, D. Q., Haron, H., Abdulazeez, A. M., and Zeeba-

ree, S. R. M. (2017). Combination of k-means clus-

tering with genetic algorithm : A review.

Zhang, Z., Tao, M., and Yuan, H. (2015). A Parking Oc-

cupancy Detection Algorithm Based on AMR Sensor.

IEEE Sensors Journal, 15(2):1261–1269.

Combining Clustering Algorithms to Extract Symmetric Clusters from Noisy Data, Applied to Parking Lots

369