Car Parking Space Detection Using YOLOv8

Muhammad Sobirin

, Tiorivaldi

and Choirul Mufit

Department of Electrical Engineering, Universitas 17 Agustus 1945 Jakarta, Jl. Sunter Permai Raya, Sunter Agung,

Tanjung Priok, Jakarta Utara, Indonesia

Department of Civil Engineering, Universitas 17 Agustus 1945 Jakarta, Jl. Sunter Permai Raya, Sunter Agung, Tanjung

Priok, Jakarta Utara, Indonesia

Keywords: Parking Lot Detection, YOLOv8, Car Detection, Available Space Detection.

Abstract: For many years, parking has been a major issue in many cities all around the world. Air pollution and traffic

congestion can be decreased by providing information about available parking spaces. Thus, the purpose of

this study is to use YOLOv8 algorithm to identify the quantity of cars and available spaces in parking lots.

Videos were captured with a camera in different scenarios at UTA'45 Jakarta, and the dataset was prepared

by extracting frames from these videos. There are no pre-labeled images in the dataset, so all of the images

have been manually annotated. Multiple object detection has been accomplished by implementing YOLOv8

algorithm to detect cars and available spaces. This paper discusses two architectures: YOLOv8 and YOLOv5.

The performance of various designs is assessed by comparing the precision, recall, and mAP values. YOLOv8

performs better than YOLOv5 when both performances are applied. In terms of mAP 0.5, mAP 0.5:0.95, and

recall, the YOLOv8 model performs better than the YOLOv5 model; the differences in the values of each

performance are 0.8%, 1.6%, and 1.2%. With a 0.5% difference in accuracy performance value, the YOLOv5

model outperforms the YOLOv8 model.

INTRODUCTION

Parking is a basic facility in every facility and

infrastructure service provider such as shopping

centers, ports, airports, etc. As time goes by the need

for parking space tends to increase with the increase

in visitors bringing private vehicles, especially cars.

Adequate parking facilities are needed for the

convenience of visitors, so parking managers provide

spacious parking spaces even with multi-level

parking patterns. This condition will become difficult

if there are only a few remaining parking slots. This

makes the driver have to search the parking lot to find

a parking lot that is still available. (Sani & Ayyasy,

2022)

Object detection or commonly called object

detection is a process used to determine the presence

of certain objects in a digital image. The detection

process can be carried out using various methods

which basically read the features of all objects in the

https://orcid.org/0009-0008-8872-0016

https://orcid.org/0000-0002-9816-573X

https://orcid.org/0000-0002-9275-1973

input image. The features of the object in the image

will be compared with the features of the reference

object or template and then compared and determining

whether the detected object is the object you want to

detect or not. (Rizkatama et al., 2021)

Artificial intelligence technology such as You

Only Look Once (YOLO) has been applied and used

in many different industries. This article presents a

smart parking system using artificial intelligence. In

complex urban environments where the number of

vehicles continues to increase, vehicle drivers have

time to find parking and traffic congestion increases

during rush hours (Acharya et al., 2018). After

entering the parking lot, it took time to find a parking

space. To alleviate these problems, a camera-based

Parking Guidance Information (PGI) system has been

investigated. (Chen & Chang, 2011).

394

Sobirin, M., Tiorivaldi, . and Muﬁt, C.

Car Parking Space Detection Using YOLOv8.

DOI: 10.5220/0012582600003821

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

In Proceedings of the 4th International Seminar and Call for Paper (ISCP UTA ’45 JAKARTA 2023), pages 394-398

ISBN: 978-989-758-691-0; ISSN: 2828-853X

METHODS

This research discusses about car parking detection

using YOLOv8. The research was carried out by

taking 5000 image of the parking space which were

checked manually one by one and classified from

morning to evening. This is because the parking space

is only busy at that time.

2.1 Original YOLO Algorithm

YOLO was introduced to the computer vision

community through a paper publication in 2016 with

titled “You Only Look Once: Unified, Real-Time

Object Detection.” (Redmon et al., 2016). The paper

reframed object detection, essentially presenting it as

a one-shot regression problem, starting with image

pixels and moving on to bounding boxes and class

probabilities. The proposed approach, based on the

concept of “unification”, allows the simultaneous

prediction of multiple bounding boxes and class

probabilities, improving both speed and accuracy.

Since its founding in 2015 until this year (2023),

the YOLO family has continued to grow at a rapid

pace. Although the original author (Joseph Redmon)

discontinued his research in the field of computer

vision with YOLO-v3 (Vidyavani et al., 2019) , the

effectiveness and potential of the main ‘unified’

concept has been expanded upon by several authors,

with the newest addition to the YOLO family coming

in the form of YOLO-v8. Figure 1 shows the

evolution timeline of YOLO.

Figure 1: Timeline of YOLO.

2.2 YOLO v8

The newest of the YOLO family was confirmed in

January 2023 with the release of YOLO-v8 (Jocher &

Tune, 2023) by Ultralytics (YOLO-v5 was also

released). Although the release of the paper is just

around the corner and many features have yet to be

added to the YOLO v8 repository, a first comparison

between the newcomer and its predecessor shows the

advantages of the new cutting-edge as his YOLO is

shown.

Figure 2 shows that when comparing YOLO-v8

with YOLO-v5 and YOLOv6 trained at image

resolution 640, all YOLO-v8 variants achieve better

throughput with a similar number of parameters,

demonstrating hardware-efficient architectural

innovations. YOLO-v8 and YOLO-v5 are introduced

by Ultralytics, YOLO-v5 offers superior real-time

performance, and based on initial benchmark results

published by Ultralytics, YOLO-v8 can focus on

Constrained Edge. There are high expectations.

Provides device provisioning with high inference

speeds.

Figure 2: YOLO-v8 comparison with predecessors (Jocher

& Tune, 2023).

The backbone part of YOLOv8 is fundamentally

the same as that of YOLOv5, and the C3 module is

supplanted by the C2f module based on the CSP

thought. The C2f module learned from the ELAN

thought in YOLOv7 and combined C3 and ELAN to

create the C2f module (Wang et al., 2022).

In the neck part, the feature fusion method used in

YOLOv8 is still PAN-FPN, which enhances the

fusion and utilization of feature layer information at

different scales. To assemble the neck module, the

creator of YOLOv8 used two up-sampling modules

and multiple her C2f modules, along with the final

separated head structure. The idea of cutting off

Car Parking Space Detection Using YOLOv8

395

YOLOx's head was used in her YOLOv8 for the last

part of the neck. We achieved a new level of accuracy

by combining confidence and regression boxes.

YOLOv8 supports all versions of YOLO and

allows you to switch between different versions at

will. It can also run on different hardware platforms

(CPU-GPU), giving you great flexibility. The

YOLOv8 network architecture diagram is shown in

Figure 3. The CBS in Figure 3 consists of a

convolution function, a batch normalization function,

and a SiLu activation function.

Figure 3: The network structure of YOLOv8.

Figure 4: SPFF and CBS structure.

Figure 5: C2f and Detect structure.

2.3 Parking Lots Detection

YOLOv8 is the latest version of YOLO by

Ultralytics. As a cutting-edge, state-of-the-art

(SOTA) model, YOLOv8 builds on the success of

previous versions, introducing new features and

improvements for enhanced performance, flexibility,

and efficiency. YOLOv8 supports a full range of

vision AI tasks, including detection, segmentation,

pose estimation, tracking, and classification as seen in

Figure 3. This versatility allows users to leverage

YOLOv8's capabilities across diverse applications

and domains.

Figure 6: YOLOv8 Ultralytics performs.

YOLOv8 Detect, Segment and Pose models pre-

trained on the COCO dataset are available, as well as

YOLOv8 Classify models pre-trained on the

ImageNet dataset. Track mode is available for all

Detect, Segment and Pose models. All models can be

seen in Table 1.

ISCP UTA ’45 JAKARTA 2023 - THE INTERNATIONAL SEMINAR AND CALL FOR PAPER (ISCP) UTA ’45 JAKARTA

396

Table 1: Models on YOLOv8.

Model

size

(pixels)

mAP



50-95

Speed

CPU ONNX

(ms)

Speed A100

TensorRT

(ms)

Params

(M)

FLOPS

(B)

YOLOv8n 640 37.3 80.4 0.99 3.2 8.7

YOLOv8s 640 44.9 128.4 1.20 11.2 28.6

YOLOv8m 640 50.2 234.7 1.83 25.9 78.9

YOLOv8l 640 52.9 375.2 2.39 43.7 165.2

YOLOv8x 640 53.9 479.1 3.53 68.2 257.8

RESULTS AND DISCUSSION

To enhance each step, the YOLOv8s algorithm was

tested and trained on the UTA’45 Jakarta Lot Parking

dataset, and its results were compared with those of

YOLOv5 to confirm that this approach could increase

the precision of space availability and car targets

detection. Lastly, in order to compare the detection

outcomes of the suggested approach and the YOLOv5

algorithm in real scenes, we chose images of

complicated sceneries in various scenarios.

It has been determined after numerous testing that

the algorithm basically iterates 438 before starting to

converge. We determined the following parameters

based on the hardware available and several testing

runs: batch size=16, epoch=2000.

3.1 Experimental Platform

Google Colab served as the online platform for the

experiments in this work, and the system hardware

included an NVIDIA T4 GPU and 16GB of RAM

(software platform: torch-2.0.1+cu118, Google

Colab).

3.2 Valuation Index

Evaluation metrics: Mean average precision (mAP),

average precision (AP), precision (P), and recall (R).

The formulas for P and R are as follows:

𝑃=

𝑇𝑃

(

𝑇𝑃+𝐹𝑃

)

(1)

𝑅=

𝑇𝑃

(

𝑇𝑃+𝐹𝑁

)

(2)

TP is the number of correctly predicted bounding

boxes, FP is the number of incorrectly judged positive

samples, and FN is the number of undetected targets.

Average Precision (AP) is the average accuracy of

the model. Mean Average Precision (mAP) is the

average value of the AP. K is the number of

categories (Lou et al., 2023). The formulas for AP and

mAP are as follows:

𝐴

𝑃=𝑝

(

𝑟

)

𝑑𝑟





(3)

𝑚𝐴𝑃 =

𝑘



𝐴

𝑃









(4)

3.3 Experimental Result Analysis

We performed the training and inference procedure

using the UTA'45 Jakarta Lot Parking dataset and

compared it with YOLOv5s to validate the detection

effect of the suggested approach on available space

and cars in this research. Students studying electrical

engineering at UTA'45 Jakarta, Indonesia, gathered

the UTA'45 Jakarta Lot Parking dataset. The dataset

was gathered from the sixth floor of the UTA'45

Jakarta building using a Logitech C270 HD Webcam.

The camera captured a variety of situations and

lighting conditions, resulting in a large number of

parking spaces and cars in complex environments.

The assessment index for this experiment is

mAP0.5, mAP0.5:0.9, which amply demonstrates the

experiment's authenticity. Table 2 displays the test

results.

Table 2: Algorithm comparison at each stage.

Detection

Algorithm

Resul

Epoch

mAP

0.5

mAP

0.5:0.95

P R

YOLOv8s 371 0.955 0.828 0.953 0.924

YOLOv5s 1237 0.947 0.812 0.958 0.912

Table 2 demonstrated that the YOLOv8s

algorithm has a certain improvement at each stage for

the detection of available spaces and cars in complex

scenarios. Furthermore, there is a 1% improvement in

the mAP 0.5 and mAP 0.5:0.95, indicating significant

Car Parking Space Detection Using YOLOv8

397

room for development. It is demonstrable that the

YOLOv8s in this experiment work quite well. The

experimental findings demonstrated that the

YOLOv8s outperforms the YOLOv5s in terms of

performance.

CONCLUSION

In contrast to conventional approaches, this study

suggests a camera sensor-based algorithm for

detecting cars and available space for smart parking

projects. It is known that the YOLOv5 and YOLOv8

models have successfully detected cars and available

spaces in parking lot images based on the research

findings that have been described. Variations exist in

parking lot detecting performance metrics. For recall,

mAP 0.5, and mAP 0.5:0.95, the YOLOv8 model

performs better than the YOLOv5 model; the

differences in the values of each performance are

0.8%, 1.6%, and 1.2%. With a 0.5% difference in

accuracy performance value, the YOLOv5 model

outperforms the YOLOv8 model.

We will keep researching camera sensors in-depth

in the future in an effort to meet our target of being

able to recognize objects in different parking lots

more accurately than current detectors as soon as

feasible.

ACKNOWLEDGEMENTS

This work was supported by UTA’45 Jakarta. The

source code for the experiments is available at the

author.

REFERENCES

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

image-based parking occupancy detection using deep

learning. CEUR Workshop Proceedings,

2087(September 2020), 33–40.

Chen, M., & Chang, T. (2011). A parking guidance and

information system based on wireless sensor network.

2011 IEEE International Conference on Information

and Automation, 601–605.

https://doi.org/10.1109/ICINFA.2011.5949065

Jocher, G., & Tune, M. (2023). Ultralytics YOLOv8.

Github.

Lou, H., Duan, X., Guo, J., Liu, H., Gu, J., Bi, L., & Chen,

H. (2023). DC-YOLOv8: Small-Size Object Detection

Algorithm Based on Camera Sensor. Electronics,

12(10), 2323.

https://doi.org/10.3390/electronics12102323

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016).

You Only Look Once: Unified, Real-Time Object

Detection. 2016 IEEE Conference on Computer Vision

and Pattern Recognition (CVPR), 779–788.

https://doi.org/10.1109/CVPR.2016.91

Rizkatama, G. N., Nugroho, A., & Suni, A. F. (2021).

Sistem Cerdas Penghitung Jumlah Mobil untuk

Mengetahui Ketersediaan Lahan Parkir berbasis Python

dan YOLO v4. Edu Komputika Journal, 8(2), 91–99.

https://doi.org/10.15294/edukomputika.v8i2.47865

Sani, A., & Ayyasy, D. H. (2022). Prototipe Deteksi

Ketersediaan Slot Parkir Berbasis Pengolahan Citra.

Journal of Applied Electrical Engineering, 6(2), 59–63.

https://doi.org/https://doi.org/10.30871/jaee.v6i2.4452

Vidyavani, A., Dheeraj, K., Reddy, M. R. M., & Kumar, K.

N. (2019). Object Detection Method Based on

YOLOv3 using Deep Learning Networks. International

Journal of Innovative Technology and Exploring

Engineering, 9(1), 1414–1417.

https://doi.org/10.35940/ijitee.A4121.119119

Wang, C.-Y., Bochkovskiy, A., & Liao, H.-Y. M. (2022).

YOLOv7: Trainable bag-of-freebies sets new state-of-

the-art for real-time object detectors.

ISCP UTA ’45 JAKARTA 2023 - THE INTERNATIONAL SEMINAR AND CALL FOR PAPER (ISCP) UTA ’45 JAKARTA

398