2024 3rd IEEE International Conference on Artificial Intelligence for Internet of Things (AIIoT 2024)

Autonomous Parking Space Detection for Electric

Vehicles Based on Advanced Custom YOLOv5
1st Deepa Balusamy 2nd Geetha Shanmugam 3rd Baazilkumar Rameshkumar
Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
Nandha Engineering College Nandha Engineering College Nandha Engineering College
Erode, India Erode, India Erode, India
[email protected] [email protected] [email protected]

4th Agathiyan Palanisamy 5th Giftson Kingston 6th Logeshraj Ravi

Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
Nandha Engineering College Nandha Engineering College Nandha Engineering College
Erode, India Erode, India Erode, India
[email protected] [email protected] [email protected]

Abstract—This research contributes to the development and devices, ensuring widespread accessibility and applicability
deployment of smart parking system utilizing cutting-edge in real-world parking environments [4].
computer vision techniques. The study utilizes a subset of the
Google Open Images Dataset focused on parking spaces to train Central to the effectiveness of the smart parking system is
a Custom Object Detection model tailored for low-computation the utilization of cutting-edge YOLOv5 architecture coupled
devices. Employing YOLOv5 architecture and the Feature with the Feature Pyramid Network (FPN). This combination
Pyramid Network (FPN), the system effectively detects parking of state-of-the-art technologies enables the system to
spaces and objects within them. Utilizing ResNetv250 as the accurately detect parking spaces and objects within them in
backbone architecture ensures precise feature extraction, real-time [5]. Additionally, the backbone architecture,
resulting in high accuracy in identifying parking spots. ResNetv250, plays a crucial role in facilitating precise feature
Experimental results reveal the efficacy of the envisioned extraction, thereby enhancing the system's overall
methodology, with Efficient YOLOv5 model achieving a mean performance and reliability. By harnessing the power of
Average Precision (mAP) of 0.970 for object detection in aerial computer vision and machine learning, the proposed smart
dimensional frames. Graphical representations of F1- parking system aims to address the challenges of Townscape
Confidence Curve and Precision-Confidence Curve parking management, offering an Elastic and efficient solution
demonstrate the model's robustness in balancing precision and for identifying and optimizing parking spaces [6].
recall, even with an imbalanced dataset. This research
underscores the potential of the proposed smart parking system II. RELATED WORKS
to revolutionize urban parking management, providing an
efficient and accurate solution for parking space identification. As the number of vehicles continues to increase, finding
parking spaces has become more challenging [7]. To address
Keywords—Smart parking system, computer vision, Custom this issue, a Smart Parking System has been developed. This
Object Detection, YOLOv5, ResNetv250, feature extraction and system includes a mobile application that allows users to
urban parking management. easily access information about available parking spaces and
efficiently manage them within a parking lot. It utilizes image
I. INTRODUCTION processing technology to recognize vehicle license plates and
The rise of urbanization and the swift growth in vehicle can automatically open and close parking lot gates as vehicles
numbers have presented notable obstacles in effectively approach. Additionally, the mobile app provides users with
managing parking availability within urban settings [1]. As details about parking availability and safety features such as
cities continue to expand, the demand for efficient parking alerts for fire and gas leaks. The core control unit of the
solutions becomes increasingly critical. Traditional methods system is a Raspberry Pi, which oversees and coordinates all
of parking management often fall short in addressing the operations. At the entrance of the parking lot, a display screen
complexities arising from limited space and escalating indicates the current availability of parking spaces. Infrared
vehicle numbers. In response to these challenges, this study proximity sensors are employed to detect the presence of
proposes a novel smart parking system that leverages vehicles at the entry gate. When a vehicle's license plate is
advanced computer vision techniques to revolutionize identified, the Raspberry Pi sends a signal to a servo motor to
parking space detection and management [2]. open the gate for a specific duration. Furthermore, when users
depart from their parking spaces, the system records the date
The proposed smart parking system relies on a Custom and time information for billing purposes.
Object Detection model trained on a curated subset of the
Google Open Images Dataset, specifically focusing on The rise of smart cities has brought technological
parking spaces. This dataset provides a diverse range of advancements, yet parking areas often lack innovation. This
images capturing various parking scenarios, including both study presents an affordable camera-based smart parking
empty and occupied parking spaces [3]. Through meticulous system for airports, utilizing inverse perspective mapping
training and optimization, the Custom Object Detection (IPM) to generate an aerial view of parking lots and guide
model is tailored to operate efficiently on low-computation drivers to available spaces [8]. Tested at Macao International
Airport, the system achieves a high accuracy rate of 97.03%

979-8-3503-7212-0/24/$31.00 ©2024 IEEE

 2024 3rd IEEE International Conference on Artificial Intelligence for Internet of Things (AIIoT 2024)
with minimal distance error. It highlights the potential of images was carefully curated to meet the specific
camera-based data for improving parking facilities in open requirements and objectives of the study. Additionally, the
spaces, offering a practical solution for smart parking dataset was downloaded in PNG format, which is a common
implementation. image file format known for its lossless compression and
compatibility with various image processing and machine
The growing need for an automated parking system to learning tools. This choice of format ensures the suitability of
enhance access and monitoring control is addressed. The
the dataset for further analysis and experimentation in the
focus lies on developing a smart parking prototype targeting context of parking lot detection research.
the management of car entrance, exit, and parking [9].
Through the integration of sensors and advanced B. Data Pre-processing
technologies, the system aims to streamline parking The dataset was obtained from Google's open-source
processes and improve efficiency. It explores features such as image repository, V6+. It comprises 10,000 images classified
vehicle occupancy detection and an Automatic License Plate into two categories: empty and occupied, each containing
Recognition System (ALPR). Validation of the ALPR's 5,000 images. Before proceeding with the analysis, pre-
performance was conducted using license plate images from processing helps to upscale the dataset into fitting shape.
Brunei, resulting in a detection rate of 97.1% and an overall These steps involved standardizing the image size to 640x640
recognition rate of 48.6%. This system is deemed suitable for pixels and applying various augmentation techniques,
real-time deployment. including rotation, cropping, and color saturation adjustments.
In crowded areas, it's tough to find parking, which leads By diversifying the images in this manner, our aim was to
to more pollution and wasted time and fuel. Using the replicate a wider range of environmental conditions and
Industrial Internet of Things (IIoT), cities can become lighting scenarios. This process enriched the dataset and
smarter, including their parking systems [10]. This paper facilitated more effective model training.
suggests a new approach called Deep Reinforcement
Learning (DRL) for smart parking systems. This system uses TABLE 1. DATA PRE-PROCESSING AND AUGMENTATION
smart cameras, fog nodes, and a cloud server. The DRL, Classes Training (70%) Validation Testing (10%)
placed in fog nodes, helps decide where cars should park by (20%)
recognizing them and finding empty spots quickly and
Occupied 4321 1538 876
accurately. Cameras at the entrance and in the parking lot
detect vehicles and send the info to fog nodes. These nodes Empty 4964 1894 915
then use a special algorithm to decide the best parking spots.
Our system is faster and more accurate than others we tested.
The augmentation process not only diversified the dataset
Enhancing parking space monitoring in rapidly expanding but also increased its size, as depicted in Table 1. After
urban areas with high vehicle density is a crucial objective. augmentation, we split the dataset into three sublines: a
We have developed a comprehensive methodology that aligns training set containing 70% of the images, a validation set
with broader urban planning goals and the evolution of smart comprising 20%, and a testing set containing the balance 10%.
cities. Our approach involves generating a specialized dataset This partitioning enables effective model training, validation,
comprising annotated images of parking spaces to determine and evaluation, ensuring that the trained models generalize
occupancy. We employ advanced techniques such as AlexNet well to unseen data. By meticulously preparing the dataset and
and ResNet-50 to extract features from these images, thereby following a rigorous experimental setup, our goal was to
improving the accuracy of parking space detection [11]. An develop accurate and dependable models for tasks such as
essential element of our methodology revolves around human, wild animal, and forest fire detection, thereby
integrating Neural Networks with various optimization contributing to advancements in computer vision and deep
algorithms including Particle Swarm Optimization (PSO), learning applications.
Grey Wolf Optimization (GWO), and Dipper Throated
Optimization (DTO). Our findings emphasize the IV. PROPOSED METHODOLOGY FOR URBAN PARKING
effectiveness of this approach in enhancing the accuracy of MANAGEMENT
parking space detection. Particularly noteworthy are the In our proposed methodology, we aim to implement a
results obtained with DTO, which demonstrate a high smart parking system using a Custom Object Detection model
accuracy rate of 0.9908. These insights contribute valuable optimized for minimal computation-based devices. This
perspectives to the ongoing conversations surrounding
model efficiently identifies parking spaces and objects within
intelligent urban planning. the parking area by leveraging advanced feature extraction
III. DATASET COLLECTION AND PRE-PROCESSING techniques and convolutional neural networks (CNNs) for
object detection. Trained on a dataset of annotated images and
A. Dataset Collection fine-tuned specifically for parking space detection, the model
The dataset used in this research was sourced from the achieves high accuracy and reliability. Its architecture is
Google Open Images Dataset, specifically version 6 or higher carefully designed to minimize computational complexity,
(v6+). This dataset is widely known for its extensive making it suitable for deployment on resource-constrained
collection of diverse images spanning various categories, devices commonly found in parking lots. Overall, our Custom
including but not limited to, animals, objects, scenes, and Object Detection model represents a significant advancement
activities [12]. For this particular study on parking lot in smart parking systems, offering a highly efficient and
detection, a subset of the dataset focusing on parking spaces accurate solution for parking space identification that holds
was selected. This subset includes images categorized into great promise for improving parking management in urban
"empty" and "occupied" classes, providing a comprehensive environments.
representation of different parking scenarios. The selection of
 2024 3rd IEEE International Conference on Artificial Intelligence for Internet of Things (AIIoT 2024)
A. Object Detection Algorithms
The YOLO model, also known as You Only Look Once,
changes how computers find objects in pictures. Instead of
checking pictures twice like before, YOLO does it all at once,
making it super quick and accurate at spotting things in real-
time. Its special way of working lets it find objects in one go,
which makes it really fast and precise [13]. This makes YOLO
great for things like self-driving cars and security cameras
where finding things quickly is important. YOLO isn't just
good at finding one thing at a time it can also find lots of
different things all at once, even in complicated situations.
This makes it really useful for computers to see and
understand the world better.
B. You Only Look Once V5
YOLOv5 is a popular model used for spotting objects in Fig. 1. Proposed YOLOv5 Model for Smart Parking System
images. Its design allows it to quickly and accurately identify
objects within visual content. Unlike earlier versions, Fig. 1 illustrates the Efficient YOLOv5 Architecture,
YOLOv5 uses a single-stage approach, making Outlooks for showcasing a novel methodology introduction in YOLOv5 S
bounding boxes and object classes directly from the entire variant. This variant utilizes the ResNetv250 as a Backbone
image in one go This improves speed and efficiency. YOLOv5 and PAnet as a Neck. This proposed architecture represents a
relies on a convolutional neural network (CNN) backbone, on significant advancement in object detection, particularly in
the EfficientDet model, to extract features from images [14]. specific environmental scenarios like identifying objects using
These features go through a series of convolutional layers and UAVs. The ResNetv250 Backbone serves as a foundational
other processes to predict bounding boxes and labels for element for feature extraction, providing a lightweight and
objects. YOLOv5 offers various versions, such as YOLOv5s, efficient backbone for processing input data. The YOLOv5
YOLOv5m, YOLOv5l, and YOLOv5x, which vary in size and Backbone ResNetv250 consists of four sets of conv layer
accuracy to suit different needs. Overall, YOLOv5 is a blocks combined with the YOLOv5 Neck and Head to identify
powerful and flexible tool for object detection in computer class predictions and bounding boxes for objects.
vision applications. Additionally, the smooth incorporation of the neck in the
proposed object detection model is highlighted.
C. Custom Object Detection Model
ResNetV250 – Backbone Architecture
YOLOv5 is a cutting-edge object tracking model
renowned for its velocity and accuracy. It employs a single- ResNetv250 is an improved version of the ResNet
stage architecture to directly predict object bounding boxes architecture, specially made for recognizing images. It's built
and various class probabilities from images. YOLOv5 offers on the foundation of ResNetv50 but with some upgrades to
flexibility to adapt to various application requirements while make it work better [17]. Like ResNetv50, it can find detailed
maintaining exceptional performance [15]. In the Efficient- features in images, which helps to accurately figure out what's
YOLOv5 Perform variant, it utilizes the ResNetv250 in a picture. ResNetv250 strikes a good balance between being
backbone, which is proficient in detecting object dimensions accurate and not needing too much computing power, so it can
in images and videos. This backbone reduces model loss and be used in lots of different situations, even where there's not a
enables the deployment of lightweight models for edge- lot of computing power available. It's well-known in the edge
cutting technologies with minimal computation on devices. of computer vision for helping us understand and analyze
images better.
The Single Shot Object Detection algorithm comprises
three essential components: Backbone, Neck, and Head.
These components are fundamental to the YOLO model. In
the Efficient YOLOv5 model, the ResNetv250 backbone is
responsible for extracting rich feature representations from
images, effectively reducing spatial resolutions and providing
crucial features for subsequent processing [16]. The Neck
model incorporates the Feature Pyramid Network (FPN) to
support backbone feature extraction and model learning. The
Head component handles object detection prediction using
multiple prediction layers. It includes prediction layers that
analyze improved feature maps to generate coordinates for
bounding boxes, scores for objectness, and probabilities for
classes. Additionally, the head applies a non-maximum
suppression (NMS) step to eliminate duplicate predictions,
resulting in a set of distinct bounding boxes accurately
depicting detected objects.

Fig. 2. ResNetV250 Architecture

 2024 3rd IEEE International Conference on Artificial Intelligence for Internet of Things (AIIoT 2024)
In Fig. 2, you can see how ResNetv250's architecture is set
up. It's based on ResNetv50 but with some improvements to
make it work even better. It uses layers that help find features
in images, and each layer has blocks that help the model learn
better [18]. These blocks have special connections that make
it easier for the model to learn and train properly, even when
it's really deep. ResNetv250 also has some clever tricks to
make it more efficient, like using certain types of layers to
keep things stable during training and using activation
functions to help it learn complex patterns. Overall,
ResNetv250 is a great choice for recognizing different things
in images, and it's used in many applications where image
recognition is important.
Feature Pyramid Network (FPN) – EYOLOv5 Neck
The Feature Pyramid Network (FPN) used in the
EYOLOv5 Neck functions by analyzing a single image and
Fig. 3. F1-Confidence Curve
generating multiple versions with varying levels of detail
[19]. It consists of two main components: the bottom-up part Fig. 3 displays the F1-Confidence Curve of Custom
and the top-down part. The bottom-up part examines the YOLOv5, which demonstrates the balance between precision
image and identifies important elements at different levels of and recall at various thresholds during training. In the case of
detail. Then, the top-down part adjusts these elements to E-YOLOv5, it achieves a mean value of 0.97 with a
match the details identified in the bottom-up analysis. By confidence threshold of 0.063, indicating the model's
combining these two components, FPN produces a pyramid performance evaluation in aerial dimensions.
of features with diverse levels of detail. This approach
enhances computer understanding of images, particularly in
tasks like detecting objects of different sizes or shapes.
In YOLOv5, the head plays a crucial role in examining
feature maps obtained from the backbone network and making
predictions about objects in the input image. Its main task is
to analyze these feature maps to identify specific features
relevant for detecting objects. After this analysis, the head
predicts the coordinates of bounding boxes, objectness scores,
and class probabilities for each detected object. Bounding box
coordinates determine where and how big the objects are,
while objectness scores indicate how likely it is that there are
meaningful objects in those boxes. Class probabilities assign
scores to different categories of objects. To enhance prediction
accuracy, the head uses non-maximum suppression to remove
redundant detections, ensuring that only the most confident
and precise predictions are kept. Fig. 4. Precision-Confidence Curve

V. EXPERIMENTAL RESULT OF CUSTOM-YOLOV5 – M Additionally, Fig. 4 illustrates the Precision-Confidence

MODEL Curve, highlighting the challenges posed by an imbalanced
The experimental result for E-YOLOv5 demonstrate its dataset where the number of negatives outweighs the
strong performance in detecting the trained classes in the positives. Despite this, the trained classes achieve a precision
packing system. The Efficient YOLOv5 achieved an value of 1.00 at a threshold of 0.981.
impressive Average Precision (mAP) of 0.970 mAP at an
intersection over union (IOU) threshold of 0.5 for objects for
the specific scenario in aerial dimensional frames. Fig. 5
showcased the Training and validation loss, along with
Performance metrices of Box loss, Class Loss, Dlf Loss,
Precision (B) and Recall graphical representation of the
efficient YOLOv5 of the proposed model across 50 epochs.
This figure highlights that backbone ResNet50-v2 reached a
Minimal loss with respect to global minima during training
and validation model.

Fig. 5. Precision-Recall Curve

 2024 3rd IEEE International Conference on Artificial Intelligence for Internet of Things (AIIoT 2024)
[2] Chen, Z., Wang, X., Zhang, W., Yao, G., Li, D., & Zeng, L. (2023).
Autonomous Parking Space Detection for Electric Vehicles Based on
Improved YOLOV5-OBB Algorithm. World Electric Vehicle
Journal, 14(10), 276.
[3] Neupane, D., Bhattarai, A., Aryal, S., Bouadjenek, M. R., Seok, U., &
Seok, J. (2024). Shine: A deep learning-based accessible parking
management system. Expert Systems with Applications, 238, 122205.
[4] Qu, S., Yang, X., Zhou, H., & Xie, Y. (2023). Improved YOLOv5-
based for small traffic sign detection under complex weather. Scientific
reports, 13(1), 16219.
[5] Xiang, P., Wen, X., Liu, Y. S., Zhang, H., Fang, Y., & Han, Z. (2023).
Retro-fpn: Retrospective feature pyramid network for point cloud
semantic segmentation. In Proceedings of the IEEE/CVF international
conference on computer vision (pp. 17826-17838).
[6] Yu, D., Guo, H., Zhao, C., Liu, X., Xu, Q., Lin, Y., & Ding, L. (2023).
An anchor-free and angle-free detector for oriented object detection
using bounding box projection. IEEE Transactions on Geoscience and
Remote Sensing.
[7] Sudhakar, M. V., Reddy, A. A., Mounika, K., Kumar, M. S., &
Fig. 6. Recall-Confidence Curve Bharani, T. (2023). Development of smart parking management
system. Materials Today: Proceedings, 80, 2794-2798.
Fig. 5 and Fig. 6 represent the Recall-Confidence Curve [8] Liu, B., Lai, H., Kan, S., & Chan, C. (2023). Camera-Based Smart
with various threshold in the trained classes with the Parking System Using Perspective Transformation. Smart Cities, 6(2),
maximum values of the trained classes as 0.97 and Precision- 1167-1184.
Recall Curve are showcase trade of between the precision and [9] Anuar, F., & Lingas, N. (2023, January). Smart campus initiative: Car
Recall of each class trained in the model of Aerial entrance, exit and parking management prototype development. In AIP
Conference Proceedings (Vol. 2643, No. 1). AIP Publishing.
Dimensional representation.
[10] Awaisi, K. S., Abbas, A., Khattak, H. A., Ahmad, A., Ali, M., &
Khalid, A. (2023). Deep reinforcement learning approach towards a
TABLE 2. PERFORMANCE COMPARISON OF OBJECT DETECTION MODELS smart parking architecture. Cluster Computing, 26(1), 255-266.
Models [email protected] Recall-confidence [11] Zaki, A. M., Towfek, S. K., Gee, W., Zhang, W., & Soliman, M. A.
curve (2023). Advancing Parking Space Surveillance using A Neural
Network Approach with Feature Extraction and Dipper Throated
Custom-YOLOv5M 0.97 0.98 Optimization Integration. Journal of Artificial Intelligence and
Metaheuristics, 6(2), 16-25.
Yolov5M 0.94 0.94
[12] Sajithra Varun, S., & Nagarajan, G. (2023). DeepAID: a design of
Yolov5l 0.96 0.97 smart animal intrusion detection and classification using deep hybrid
neural networks. Soft Computing, 1-12.
[13] Kaur, R., & Singh, S. (2023). A comprehensive review of object
Table 2 represents the performance comparison of various detection with deep learning. Digital Signal Processing, 132, 103812.
yolov5 models and its parameters accuracy value (mAP) are [14] Nguyen, D. L., Vo, X. T., Priadana, A., & Jo, K. H. (2023, February).
represented by an object mean average precision score to YOLO5PKLot: A Parking Lot Detection Network Based on Improved
YOLOv5 for Smart Parking Management System. In International
compare the object detection accuracy in particular Workshop on Frontiers of Computer Vision (pp. 95-106). Singapore:
environments. Springer Nature Singapore.
[15] Bharathi, V., Vishwaa, M., Elangkavi, K., & Krishnan, V. H. (2023).
VI. CONCLUSION A CUSTOM YOLOV5-BASED REAL-TIME FIRE DETECTION
The proposed smart parking system offers a promising SYSTEM: A DEEP LEARNING APPROACH. Journal of Data
Acquisition and Processing, 38(2), 441.
solution to the challenges of urban parking management. By
leveraging advanced technologies such as the Custom Object [16] Li, M., Pi, D., & Qin, S. (2023). An efficient single shot detector with
weight-based feature fusion for small object detection. Scientific
Detection model, YOLOv5, Feature Pyramid Network Reports, 13(1), 9883.
(FPN), and ResNetv250, the system demonstrates its ability [17] Shi, Q., Yang, Y., Sun, Z., & Lee, C. (2022). Progress of advanced
to accurately detect and manage parking spaces in real-time. devices and internet of things systems as enabling technologies for
With its capability to operate efficiently on low-computation smart homes and health care. ACS Materials Au, 2(4), 394-435.
devices, the system ensures accessibility and applicability in [18] Hindarto, D. (2023). Use ResNet50V2 Deep Learning Model to
various parking environments. By improving the efficiency Classify Five Animal Species. Jurnal JTIK (Jurnal Teknologi
of parking space detection and management, our smart Informasi dan Komunikasi), 7(4), 758-768.
parking system contributes to alleviating congestion and [19] P. T, R. Thangaraj, P. P, U. R. M and B. Vadivelu, "Real-Time
Handgun Detection in Surveillance Videos based on Deep Learning
enhancing overall urban mobility. With further refinement Approach," 2022 International Conference on Applied Artificial
and implementation, this system holds great potential to Intelligence and Computing (ICAAIC), Salem, India, 2022, pp. 689-
revolutionize parking management and enhance the quality 693, doi: 10.1109/ICAAIC53929.2022.9793288.
of life in urban areas.
REFERENCES
[1] Chen, Z., Wang, X., Zhang, W., Yao, G., Li, D., & Zeng, L. (2023).
Autonomous Parking Space Detection for Electric Vehicles Based on
Improved YOLOV5-OBB Algorithm. World Electric Vehicle
Journal, 14(10), 276.