Article Not peer-reviewed version

Evaluating the Performance of

Ensembled YOLOv8 Variants in Smart
Parking Applications for Vehicle
Detection and License Plate
Recognition under Varying Lighting
Conditions

Ripunjay Singh * , Sarthak Goyal , Shivam Agarwal , Subho Upadhyay

Posted Date: 18 April 2024

doi: 10.20944/preprints202404.1216.v1

Keywords: YOLOv8; Smart Parking; TOPSIS; Object Detection; YOLOv8 ensembling

Preprints.org is a free multidiscipline platform providing preprint service that

is dedicated to making early versions of research outputs permanently
available and citable. Preprints posted at Preprints.org appear in Web of
Science, Crossref, Google Scholar, Scilit, Europe PMC.

Copyright: This is an open access article distributed under the Creative Commons
Attribution License which permits unrestricted use, distribution, and reproduction in any
medium, provided the original work is properly cited.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

Disclaimer/Publisher’s Note: The statements, opinions, and data contained in all publications are solely those of the individual author(s) and
contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting
from any ideas, methods, instructions, or products referred to in the content.

Article
Evaluating the Performance of Ensembled YOLOv8
Variants in Smart Parking Applications for Vehicle
Detection and License Plate Recognition under
Varying Lighting Conditions
Ripunjay Singh *, Sarthak Goyal, Shivam Agarwal and Subho Upadhyay
Department of Electrical, Dayalbagh Educational Institute (deemed to be university), Dayalbagh Road, Agra,
282005, Uttar Pradesh, India; [email protected] (S.G.); [email protected] (S.A.);
[email protected] (S.U.)
* Correspondence: [email protected]

Abstract: This paper investigates the efficacy of YOLOv8 variants for vehicle detection and license
plate detection within smart parking applications, emphasizing performance under varying ambient
lighting conditions. The proposed system is to seize full video frames, extracts regions of interest
containing vehicles, and feeds them into separate, pre-trained YOLOv8 models – one dedicated
to vehicle detection and another for license plate detection. Four YOLOv8 variants, nano, small,
medium, and large, are evaluated. As a pre-processing step, the images are processed with the help of
OpenCV and Pillow libraries to adjust the luminosity and increase the images’ DPI so that they would
be easy to perceive by the Tesseract OCR engine. Sixteen potential combinations arise from pairing
the four YOLOv8 models for vehicle and license plate detection tasks. To identify the most suitable
combinations, we employ the Multi-Criteria Decision-Making method, specifically Technique for
Order of Preference by Similarity to Ideal Solution (TOPSIS) analysis. This analysis considers 4 critical
metrics: Precision, mean average precision at 95% Intersection of Union threshold, Recall and Total
Inference Time. The objective is to achieve an optimal balance between high accuracy and real-time
processing. Following the selection of optimal YOLOv8 combinations through TOPSIS analysis, we
assess their performance under varying ambient light intensity (measured in lux). This evaluation
aims to identify the most robust model combinations that ensure accurate vehicle and license plate
recognition across the diverse lighting conditions encountered in real-world environment.

Keywords: YOLOv8; Smart Parking; TOPSIS; Object Detection; YOLOv8 ensembling

1. Introduction
The remarkable economic progress and population growth in megacities are a foothold for
one critical challenge: an increasing number of cars and no parking space. The severe imbalance
between the parking supply and demand cascade into numerous constraints of urban mobility. Traffic
congestion is one of the persisting problems of contemporary metropolises where vehicles have to
spend excessive time to find a parking site. According to the study conducted by INRIX and cited in
the article in USA Today, the drivers may waste 17-107 hours a year in search for parking sites [1]. It
is calculated that a US driver spends approximately $345-2,243 on fuel, emissions and time loss in
search for parking. In sum up, the estimated price for the USA is annually $73 billion [2]. It manifests
in vehicles’ gridlock, lost man-days due to inefficiency, and decrease of the general quality of life.
This study proposes the implementation of a Smart Parking Management System (SPMS) as a
technology-driven solution to address these critical challenges. The SPMS would leverage cutting-edge
object detection AI models to ensure efficient parking detection and functionality. This research
endeavor will analyze the system’s performance with respect to efficiency, accuracy, and its ability to
function effectively in real-world conditions, including variations in ambient lighting.

© 2024 by the author(s). Distributed under a Creative Commons CC BY license.

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

2 of 25

Overall, this study will compare the proposed system’s performance in the above mentioned
axes. The multifaceted solution will help to ensure the ease of traffic management by reducing the
search time and increasing the capacity of the used parking space. Moreover, the money obtained
from parking services may be an additional source of revenue for the city authorities to invest in new
parking spaces.

2. Literature Review
This section reviews recent advancements in smart parking management systems, focusing on
automated parking time violation tracking and vehicle detection algorithms.
Several studies propose innovative solutions to address parking challenges. One approach
utilizes YOLOv8 [3] for accurate vehicle detection, coupled with DeepSORT [4] [5] or OC-SORT [6] for
robust multi-object tracking. DeepSORT exhibits exceptional accuracy across diverse datasets, while
OC-SORT excels in scenarios with minimal variations. These findings provide valuable insights for
selecting algorithms based on specific parking lot requirements. Future research directions include
customizing training data to enhance detection accuracy and exploring synchronization of algorithms
across multiple cameras for comprehensive monitoring of large parking areas [7]. Additionally,
advancements in detection architectures, such as the improved YOLOv5 [8], demonstrate promising
performance with high mean Average Precision (mAP), facilitating real-time deployments [9]. Further
research is recommended to optimize detection for smaller vehicles and compare the efficiency of
YOLOv5 with YOLOv8 for optimal selection.
Implementing smart parking systems offers numerous benefits, as it may potentially decrease the
ecological problems in parking and areas adjacent to them, since it helps decrease vehicle ‘ search time.
In addition, it deals with a load of vehicle emission. However, the development of future studies is
relevant and may focus on their customization and diversified data as well as across-multiple cameras
notification [10].
This review highlights the significant progress in smart parking management systems, particularly
in utilizing advanced detection and tracking algorithms. These advancements pave the way for
cost-effective, scalable solutions with the potential to generate substantial economic and environmental
benefits. Future research should focus on customization, multi-camera integration, and robust training
data to further optimize these systems for real-world applications [11].

3. Limitations of Existing Research

One of the most important and urgent problems in modern research activities and product
design is the lack of research focused on the relationship between model complexity, pre-processing
techniques, and their influence on the performance of the designed system, especially in devices with
scant resources for deployment. While smart parking systems based on YOLO [12] object detection
is a significant achievement, there is a current research gap regarding the trade-off between model
complexity; the use of some pre-processing techniques, like luminosity balancing; and the performance
of the system. The existing sources cover on limited exploration of the influence of YOLOv8 variants
on smart parking systems; the problem of the influence of pre-processing methods on license plate
recognition ; and a limited scope of lighting variations in the research.

3.1. Research Gaps

3.1.1. Limited Exploration of YOLOv8 Variant Trade-offs

While the problem of the use of YOLOv8 variants in smart parking has been every time studied
from the perspective of its accuracy in terms of vehicle detection, other variants of YOLOv8, like nano,
have been not researched. A complete evaluation of all YOLOv8 variants in terms of their application
in resource-limited smart parking systems is needed to define the optimal model in terms of a trade-off
between accuracy and inference speed.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

3 of 25

3.1.2. Uncertain Impact of Pre-processing on License Plate Detection

While some researchers state that there was a stated need for the use of pre-processing techniques,
like luminosity balancing, applied to YOLOv8 in terms of plate recognition, there is a limited
exploration of what is the influence of its use on License Plate Detection performance. An investigation
into these relationships is critically needed for the purpose of optimizing the pre-processing pipeline.

3.1.3. Limited Scope of Lighting Variation

The conduction of the experiments below has the limitation of the absence of the study of YOLOv8
variants application across a broad range of lighting variations. No mobile hardware specifically
designed for a certain smart parking model was used.

3.2. Bridging Gaps

3.2.1. Evaluating All YOLOv8 Variants:

Multiple Criteria Decision Making techniques will be used to evaluate all YOLOv8 variants, like
nano, small, medium, large, less available than small or typically-used nano, and vise versa, in terms of
their application for vehicle detection and LPR. The result of the research will allow the identification
of the optimal model and its variant that can be used in resource-limited smart parking.

3.2.2. Applying Pre-processing techniques for OCR enhancement

The impact of pre-processing techniques, such as luminosity balancing [13] and DPI enhancement
[13], on improving the accuracy of OCR will be assessed.

3.2.3. Analyzing Performance under Varying Lighting Conditions

The most optimal YOLOv8 combinations, identified through Multi-Criteria Decision Making
(MCDM) techniques, will be further evaluated under various lighting intensities (lux). This will ensure
the system’s efficacy and robustness across diverse real-world parking environments.
By addressing these research gaps, this study will contribute to the development of more efficient
and adaptable smart parking systems that are optimized for resource-constrained devices and perform
effectively under a wider range of lighting conditions, ultimately leading to a more robust and practical
implementation of smart parking solutions.

4. Methodology
As shown in Figure 1, the execution of the proposed approach requires a robust methodology in
order to facilitate the required functioning.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

4 of 25

Figure 1. The proposed method facilitates scalable street parking management by leveraging
underutilized road shoulders.

The task of this research is to conduct a profound analysis of different configurations of the
YOLOv8 models to optimize vehicle and license plate recognition in the smart parking system. Herein,
to ensure maximum feasibility of the project, several vital stages of work have been followed, with
each stage being organized in accordance with the core principles of cogent research.

4.1. Datasets Acquisition

4.1.1. Vehicle Dataset [14]:

The car detection model undergoes training using an extensive dataset comprising 15,322 images
annotated in the YOLOv8 format. This format provides essential bounding boxes and class labels for
each object within the images, enabling the model to effectively discern the unique visual attributes
of cars. To ensure consistent image representation and optimize training efficiency, we implement a
pre-processing pipeline.
This pre-processing pipeline consists of several critical steps:

• EXIF Data Removal and Auto-Orientation: Initially, the image pixel data undergoes automatic
orientation correction based on embedded EXIF information. Subsequently,the EXIF data itself is
removed to ensure that the model focuses solely on the image content, minimizing the potential
inconsistencies.
• Uniform Resizing (Stretch): All images are uniformly resized to a standardized dimension of
640x640 pixels using a stretching approach. This standardization ensures compatibility with the
YOLOv8 model architecture and facilitates efficient processing during the training phase.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

5 of 25

• Data Augmentation for Enhanced Generalizability: In this project, following data augmentation
techniques were utilized on each source image, resulting in the creation of three additional
augmented variations:

- Random Cropping (0-15%): This technique simulates potential variations in object

positioning within the frame by randomly cropping a small portion (0-15%) of the image
from various locations.
- Random Brightness Adjustment (+/- 9%): Controlled variations in lighting conditions are
introduced through brightness adjustments of +/- 9%, enhancing the model’s performance
in real-world scenarios with diverse lighting environments.
- Random Gaussian Blur (0-2.5 Kernel Size): To simulate slight out-of-focus effects
encountered during real-world car detection tasks, a random Gaussian blur with a kernel
size varying between 0 and 2.5 pixels is applied.

Following the comprehensive pre-processing pipeline, the dataset undergoes division into
training (88%), testing (8%), and validation (4%) sets.

4.1.2. License Plate Dataset [15]:

This project relies on a specialized dataset comprising 12,884 images for training and testing
purposes. These images encompass feature annotated license plates in the YOLOv8 format. Such
annotation facilitates the YOLOv8 models in discerning the distinct visual attributes of license plates
from the background.
The pre-processing pipeline for this dataset comprises the following steps:
• EXIF-based Auto-Orientation and Removal: Automatic orientation correction of image pixel data
is performed based on embedded EXIF metadata. This guarantees consistent image presentation
and eliminates potential discrepancies arising from varied camera orientations.
• Uniform Resizing (Stretch): All images are uniformly resized to a standardized dimension of
640x640 pixels using a stretching method. This standardization ensures compliance with the
input requisites of the YOLOv8 model architecture and enhances processing efficiency during
training.
• Data Augmentation for Enhanced Generalizability: Data augmentation techniques are
employed to expand the dataset artificially and bolster the robustness of the YOLOv8 models.
This involves generating additional variations of each source image to increase training data
diversity. Three specific random augmentation techniques are applied to create three additional
augmented versions for each image:

- Random Cropping (0-30%): A random portion (0-30% of the total image area) is cropped
from various locations within the image, encouraging the model to focus on central objects
and relevant image regions.
- Random Brightness Adjustment (±21%): Image brightness is randomly adjusted by altering
pixel intensity values within a range of -21% to +21% relative to the original image,
broadening the model’s adaptability to diverse lighting environments.
- Random Gaussian Blur (0-2.5 pixels): A Gaussian blur filter with a randomly chosen kernel
size between 0 and 2.5 pixels is applied to the image, enhancing the model’s ability to handle
variations in image quality during real-world deployments.

Following pre-processing and augmentation, the dataset is divided into training, testing, and
validation sets. An 82% serves as the training set, while the remaining 18% is split into a 16% test set
and a 2% validation set.

4.2. YOLOv8 Model Selection and Training Regimen

The next part of our study covers the selection, as well as the training, of four types of YOLOv8,
which are nano, small, medium, and large. All four variants receive individual training using a relevant
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

6 of 25

dataset. In particular, the vehicle dataset is used to provide training for the vehicle detection model,
and the license plate model is trained with the help of a license plate dataset.

4.3. System Design and Pre-processing Techniques

At the core of our approach is development of a sophisticated system architecture, which involve
a two-stage object detection system: The former encompasses an extracted Region of Interest on which
a vehicle is detected. Subsequently, the extracted ROI then pass through a license plate recognition
model. The detected license plate is then subject to several preprocessing steps, to further improve the
accuracy of the Tesseract OCR Engine [16]. It is worth noting that OpenCV [17] library facilitates the
luminosity balancing, which allows for consistent lighting condition. Additionally, the resolution of
license plates images is improve, through the resizing of images in a DPI setting, which is facilitated
the Pillow [18] library.

4.4. Model Evaluation and TOPSIS-driven Selection

From the perspective of the adopted research paradigm, it could be observed that a crucial focus
is placed on conducting a reliable evaluation of every YOLOv8 combination. In this context, there are
16 potential configurations with regard to such combinations, meaning that each model that is derived
from the addition of the four vehicle detection models to the four license plate detection models must
be analyzed. To carry out the evaluation, a broad range of diverse evaluation metrics is used, and it
includes metrics such as Precision, mAP50-95, Recall and inference speed. Interestingly, the Technique
for Order Preference by Similarity to Ideal Solution, or TOPSIS, is utilized in order to determine the
best YOLOv8 combination. It is an established method within the field of Multi-Criteria Decision
Making, and it helps to achieve a final result while taking into consideration various possible options.
Overall, through the application of such an approach, it is possible to establish an effective evaluation
process that is well-balanced in terms of such critical properties as accuracy and inference speed due
to the use of predefined weightings.

4.5. Performance Analysis under Varying Lighting Conditions

After conducting a TOPSIS analysis of the robustness of the top YOLOv8 combinations, it was
decided that those combinations would be put to the test in an environment with controlled lighting
while varying the levels of ambient light intensity. With images of both vehicles and license plates taken
in the described conditions and precise lux meter readings, it is hoped to understand how sensitive the
results are to the lighting levels and changes. The data from the lux meter will be combined with the
information extracted from the class confidence scores of the performance in vehicle detection and
license plate detection to examine the robustness of the best combination.

4.6. Model Training:

Leveraging the computational power of a P100 GPU on the Kaggle platform, YOLOv8 models
were trained with a focus on achieving an optimal balance between training efficiency and model
performance. This involved the selection of key hyperparameters:

• Epochs (50): A total of 50 epochs were chosen to ensure comprehensive exposure of the models
to the training data. Validation performance was rigorously monitored throughout training to
prevent overfitting and determine the most suitable number of epochs.
• Image Size (640x640): A standardized image size of 640x640 pixels was selected for all input
images. This choice balances the ability to capture essential image information with maintaining
computational efficiency during training.
• Training Batch Size (16): The training batch size was set to 16 images per iteration, considering
both GPU memory utilization and the benefits of gradient updates.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

7 of 25

• Nominal Batch Size (64): Specific to the YOLOv8 architecture, a nominal batch size of 64 was
employed to leverage gradient accumulation across multiple training batches, enhancing training
stability and convergence.

the rigrous hyperparameter tuning process underscores the commitment to achieving efficient
traing that fostes superior model performance on target tasks within our smart parking application,
while also promoting model generalizability for real-world scenarios

Figure 2. Confusion Matrix for YOLOv8m for Vehicle Detection

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

8 of 25

Figure 3. Confusion Matrix for YOLOv8s for Vehicle Detection.

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

9 of 25

Figure 4. Confusion Matrix for YOLOv8m for License Plate Detection.

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

10 of 25

(a) Plot showing variation in Precision (Equation (4)) per epoch.

(b) Plot showing variation in mAP50-95 (Equation (7)) per epoch.

Figure 5. Training Insights of Precision (Equation (4)) and mAP50-95 (Equation (7)) for Top 2
combinations according to TOPSIS analysis (Table 2).
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

11 of 25

Figure 6. System flow of the proposed design.

5. Details of Approach

5.1. Extraction of Region of Interest:

A simple functionality of the OpenCV [17] library was created to define and export regions of
interest from the full frames. The module allows the user to select an area of interest on the video
by using the mouse. After the left mouse button is clicked, the module captures the coordinates of
the mouse for the top-left corner of the selected area. When the left mouse button is released, the
coordinates for the bottom-right corner are captured. Then, the module draws a rectangle for the
defined region on the video. User can save the selected region by pressing any key function. This
rudimentary functionality allows the user to capture desired portions of the video for further analysis.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

12 of 25

Figure 7. Algorithm 1 enables users to define a region of interest (ROI) by dragging the cursor from
point A to point B (or vice versa).

Algorithm 1 ROI Extraction using OpenCV

1: Step1: Import necessary libraries
2: - OpenCV
3: Step 2: Define global variables:
4: - frame: current video frame
5: - roi_start: starting coordinates of the ROI
6: - roi_end: ending coordinates of the ROI
7: Step 3: Initialize the video capture object to read frames from the video source.
8: Step 4: Define a function to handle mouse events:
9: function ON M OUSE(event, x, y, flags, param)
10: if event is left mouse button down then
11: Capture starting coordinates of ROI (x, y)
12: end if
13: if event is left mouse button up then
14: Capture ending coordinates of ROI (x, y)
15: Draw rectangle on the frame to visualize the ROI
16: Display the frame with ROI
17: end if
18: end function
19: Step 5: Create a named window for displaying the video frame.
20: Step 6: Register the mouse callback function to the named window created in step 5.
21: Step 7: Loop over frames:
22: while video is not ended do
23: Read the next frame from the video capture object.
24: Display the frame in the named window.
25: end while
26: Step 8: Check for key press events:
27: if key is pressed then
28: if key is designated key for saving ROI then
29: Save the defined ROI as an image file.
30: end if
31: end if
32: Step 9: Release the video capture object and close all OpenCV windows.

5.2. YOLO Algorithm

The You Only Look Once algorithm [12] is one of the key technologies in the field of computer
vision, as it is a single-stage method for object detection. While multi-stage algorithms require iterative
calculations, YOLO reduces them thanks to a deep convolutional neural network(Figure 8) predicting
bounding boxes and class probabilities for objects in an image. These relatively streamlined calculations
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

13 of 25

significantly increase the processing speed, which can be invaluable in real-time applications. The
calculation results are based on a grid system in which each cell outputs bounding boxes and the
probability that this cell has an object. Several proposed bounding boxes and related probability values
are calculated. Non-Maximum Suppression removes all boxes except the one with the best probability.

Figure 8. Overview of YOLO architecture [12].

The next step in the algorithm is the configuration of the model for the purposes of detecting. Ren
et al. showed that performance improvement is possible if extra convolutional and connected layers
are added [19]. The extra two layers in the previous example are not sufficient, so the given researchers
added four more convolutional layers and two more fully connected layers to the model, initializing
their weights randomly. The input resolution of the network is important to increase the amount of
visual data received by the program being detected. To calculate its resolution, the size of the original
image is 224×224 pixels, in this case, the resolution is 448×448 pixels. The final layer of the model in
question has to predict both class probabilities and the coordinates of the bounding box. To normalize
all models, the width and height of the bounding box are also within the image, so they are between 0
and 1. In addition, the x and y coordinates of the bounding box are the coefficient’s offsets relative
to the center of a certain cell of the grid, ranging from 0 to 1. The last layer uses a linear activation
function, all other layers use a leaky rectified linear activation [12]:
(
x, if x > 0
ϕ( x ) = (1)
0.1x, otherwise

During training, it is desired for each object to be responsible for predicting only one bounding
box. For this purpose, one predictor will be assigned to be responsible for predicting the object based
on which prediction has the highest. Separation of responsibilities contributes to further training of
predictive bounding boxes in various parameters such as the size or aspect of an object it becomes
expert in predicting. In the given case, the score for the given algorithm is reduced to facilitate training.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

14 of 25

The formula for calculating the loss of the multi-part function [12] being optimized during the training
is given as:
S2 B h i
λcoord ∑ ∑ 1ij ( xi − x̂i )2 + (yi − ŷi )2
obj

i =0 j =0
S2
" q 2 #
B √ 2 p
∑∑
obj p
+λcoord 1ij wi − ŵi + hi − ĥi
i =0 j =0
(2)
S2 B
2 S2 B
2
+∑ ∑ 1ij ∑ ∑ 1ij
obj noobj
Ci − Ĉi + λnoobj Ci − Ĉi
i =0 j =0 i =0 j =0
S2
+ ∑ 1i ∑
obj
( pi (c) − p̂i (c))2
i =0 c∈ classes
obj obj
where 1i denotes if object appears in cell i and 1ij denotes that the jth bounding box predictor in
cell i is "responsible" for the corresponding prediction.
The You Only Look Once algorithm has certain advantages such as the very fast processing of
real-time images and the rather simple architecture (Figure 8) compared to the multistage detection
algorithms. At the same time, it has a series of disadvantages such as potential trade-offs in accuracy
and certain difficulties in parsing smaller or a large number of objects. These challenges have led to
the modification of the YOLO algorithm, which have resulted in the subsequent development of the
YOLOv2, YOLOv3, YOLOv4, YOLOv5, YOLOv6, YOLOv7, YOLOv8 and recently released YOLOv9
models.

5.3. Tesseract OCR

In this study focusing on evaluating YOLOv8 variants in smart parking applications, the
incorporation of Tesseract [16] is key within the OCR domain. Tesseract’s approach to OCR processing
follows a systematic pipeline that incorporates connected component analysis at its initial stages,
which comes in handy in managing scenarios with inverse text. The subsequent steps in the pipeline
incorporate outline grouping, text line organization, character spacing analysis, and a two-pass
recognition process, which culminate in fuzzy space resolution and x-height hypothesis check.
Tesseract’s journey from a research manuscript to being one of the leading OCR engine is in line
with the existing computer vision spirit of cooperation.
In this study, we use PyTesseract [20], which is the Python wrapper for the Tesseract OCR
engine, to unify Python and Tesseract. The incorporation of Tesseract in our pipeline is critical in
ensuring it is easily integrated into Python, thus linking developers with a variety of functionalities.
The functionalities demonstrated among others include the loading and pre-processing of an image
containing texts, the recognition of the loaded image using Tesseract, the extraction of the recognized
text, and customization of the Tesseract’s parameters, specifically the language feature in assessing
texts from different languages. This utilization of the PyTesseract indicates the high level of satisfaction
concerning the YOLOv8 variants for vehicle detection and license plate recognition in smart parking
applications under different lighting conditions. Additional alterations have also been executed to the
identified enhancement, including the Luminosity Modulation and DPI enhancement.

5.4. Luminosity Modulation

Luminosity modulation is highly beneficial in Optical Character Recognition (OCR), and
specifically, for license plate recognition. It allows the adjustment of brightness and contrast, thus
making the images clearer and the taught characters more distinguishable from the background.
Consequently, the use of luminosity modulation because allows improving the detection process as a
whole. On a similar note, the capacity to adjust luminosity is particularly suitable for nighttime and
in undertones. Notably, it enhances the applicability of OCR in poor lighting conditions, making the
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

15 of 25

features of the examined objects, in this case, license plates, more discernible. Moreover, it should be
noted that luminosity modulation has a positive effect in reducing signal noise . The latter effect is
essential since it creates premises for using the OCR feature in the plates’ detection. The role played
by luminosity modulation features in the moistening the glare in the photos also helps enhance the
descriptiveness of the photo. Particularly, substantial benefits can be experienced during the process
of photodetector interference calibration. As a result, luminosity modulation can be deemed an
enhancement that improves the sustainability and effectiveness of the OCR in any environment. Most
importantly, it also raises the accuracy of the OCR, therefore, improving a range of applications, such
as automatic toll monitoring, car tracking, and so forth.
Luminosity modulation (Figure 9) is implemented with the help of OpenCV [17] library according
to the following equation [13]:

Luminosity = 0.2126 × Red Component

+0.7152 × Green Component (3)
+0.0722 × Blue Component

(a) Before Luminosity Adjustments

(b) After Luminosity Adjustments

Figure 9. Effect of Luminosity Adjustment.

5.5. DPI Enhancements

The increase in image resolution, measured in dots per inch, improves the performance of Optical
Character Recognition for license plate applications in a demonstrable way. This happens because
higher resolution images have more detailed and clearer purposes providing more data for analysis. As
a result, the OCR algorithms have an easier time distinguishing different characters. The application of
this fact in practice resulted in a meaningful reduction of recognition errors and a subsequent increase
in the accuracy of license plate detection and extraction. Another significant benefit of heightened DPI
is that it improves performance in scenarios that are problematic even for human analysts. Low-light
conditions, as well as an image that is partially obscured, damaged, or distorted in some other way,
contain more data that is useful for the OCR algorithms. This allows one to improve the overall
performance of the system. We increased the DPI using the Pillow [18] library, bringing the DPI to the
constant level of 600.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

16 of 25

6. Results
In this section, the evaluation metrics utilized to comprehensively evaluate the performance of
the YOLOv8 models for the car detection and license plate recognition within our smart parking
application are thoroughly examined. These metrics shed light on the accuracy, completeness, and
overall effectiveness of the models in fulfilling their designated tasks.

6.1. Precision:
Precision gauges the accuracy with which the model identifies objects such as cars or license
plates. It represents the proportion of correctly identified objects among all detections reported by the
model. A high precision value indicates a strong correlation between the model’s positive detections
and the actual objects present in the scene. Mathematically, it is defined as:
True Positive (TP)
Precision = (4)
True Positive (TP) + False Positive (FP)
where True Positive (TP) represents the number of accurately detected objects by the model, and False
Positive (FP) denotes the instances of incorrectly identified objects (e.g., non-car objects misclassified
as cars).

6.2. Recall:
Recall complements precision by measuring the comprehensiveness of the model’s detections. It
reflects the proportion of actual objects (cars or license plates) successfully identified by the model.
The recall value is calculated as:
True Positive (TP)
Recall = (5)
True Positive (TP) + False Negative (FN)

Here, False Negative (FN) represents the number of actual objects that the model fails to detect.
A high recall value indicates that the model effectively identifies a significant portion of the existing
objects within the image.

6.3. Intersection Over Union (IoU):

IoU (Figure 10) quantifies the area of overlap between the predicted and ground truth bounding
boxes relative to their total area. It is computed as:

Area of Overlap
IoU = (6)
Area of Union
It plays a fundamental role in evaluating the accuracy of object localization.

Figure 10. A pictorial representation of various performance metrics [21].

6.4. Mean Average Precision (mAP):

mAP provides a comprehensive evaluation by averaging the precision values obtained at various
Intersection over Union (IoU) thresholds. IoU measures the overlap between the predicted bounding
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

17 of 25

box and the ground truth bounding box (actual object location). In this study, mAP50 as well as
mAP50-95 is reported, that is calculation of mAP with an IoU threshold of 0.5 and 0.95 respectively,
indicating at least 50% (for mAP50) or 95% (for mAP50-95) overlap is necessary between Ground Truth
and Prediction, for a detection to be considered accurate. The formula for Mean Average Precision
(mAP) is given by:

N
1
mAP =
N ∑ APi (7)
i =1

where, mAP is the Mean Average Precision, N is the number of classes, and APi is the Average
Precision for class i.

6.5. F1-Score:
The F1-Score offers a balanced evaluation by calculating the harmonic mean of precision (Equation
(4)) and recall (Equation (5)). This metric provides a robust assessment of model performance, less
susceptible to outliers compared to individual precision and recall values. It is expressed as:

2 × Precision × Recall
F1 − Score = (8)
Precision + Recall
Valuable insights into the strengths and weakness of the YOLOv8 models for the vehicle and
license plate detection tasks within our smart parking system are gained through and in-depth analysis
of these evaluation metrics. This analysis facilitates informed decision-making regarding the most
suitable YOLOv8 variant for real-world application deployment.
In Table 1, a comprehensive comparison of a performance metrics for various YOLOv8 variants
for vehicle detection and license plate detection respectively, is presented. The evaluation encompasses
precision (Equation (4)), recall (Equation (5)), mean Average Precision (mAP) (Equation (7)) at 50%
and 95% IoU (Equation (6)) thresholds, and F1 scores (Equation (8)).

Table 1. Comparison of metrics for Vehicle and License Plate Identification of different YOLO Models

Vehicle Detection License Plate Detection

Model Precision Recall mAP50 mAP50-95 F1 Score Precision Recall mAP50 mAP50-95 F1 Score
Nano 0.847 0.840 0.815 0.719 0.843 0.949 0.848 0.916 0.711 0.896
Small 0.863 0.843 0.821 0.747 0.852 0.942 0.853 0.916 0.718 0.895
Medium 0.864 0.849 0.853 0.768 0.856 0.942 0.86 0.919 0.716 0.899
Large 0.874 0.854 0.849 0.787 0.864 0.946 0.859 0.919 0.721 0.900
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

18 of 25

(a) Comparison of metrics for vehicle detection.

(b) Comparison of metrics for license plate detection.

Figure 11. A comparative graphical representation of various metrics of Table 1.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

19 of 25

7. Analysis

7.1. TOPSIS Analysis:

The Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) [22,23] is a method
for Multiple Criteria Decision Making (MCDM), initially conceived by Ching-Lai Hwang and Yoon in
1981, with subsequent refinements by Yoon in 1987 and Hwang, Lai, and Liu in 1993. TOPSIS operates
on the principle that the preferred alternative should be closest to the positive ideal solution (PIS) and
farthest from the negative ideal solution (NIS) in geometric distance.

7.1.1. Procedure
The TOPSIS process is carried out as follows:

• Step 1: Create an evaluation matrix consisting of m alternatives and n criteria, with the intersection
of each alternative and criteria given as xij . Therefore, we have a matrix ( xij )m×n .
• Step 2: The matrix ( xij )m×n is then normalized to form the matrix

R = (rij )m×n , (9)

using the normalization method

xij
rij = q ,
∑m 2
k =1 xkj (10)
i = 1, 2, . . . , m, j = 1, 2, . . . , n.

• Step 3: Calculate the weighted normalized decision matrix

tij = rij · w j , i = 1, 2, . . . , m, j = 1, 2, . . . , n (11)

.
where w j = Wj ∑nk=1 Wk , j = 1, 2, . . . , n so that ∑in=1 wi = 1, and Wj is the original weight given
to the indicator v j , j = 1, 2, . . . , n.
• Step 4: Determine the worst alternative ( Aw ) and the best alternative ( Ab ):

Aw = ⟨max(tij | i = 1, 2, . . . , m) | j ∈ J− ⟩,
⟨min(tij | i = 1, 2, . . . , m) | j ∈ J+ ⟩
≡ {twj | j = 1, 2, . . . , n},

Ab = ⟨min(tij | i = 1, 2, . . . , m) | j ∈ J− ⟩,
⟨max(tij | i = 1, 2, . . . , m) | j ∈ J+ ⟩
≡ {tbj | j = 1, 2, . . . , n},

where,

J+ = { j = 1, 2, . . . , n | j}

associated with the criteria having a positive impact, and

J− = { j = 1, 2, . . . , n | j}

associated with the criteria having a negative impact.

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

20 of 25

• Step 5: Calculate the L2 -distance between the target alternative i and the worst condition Aw
v
u n
= t ∑ (tij − twj )2 ,
u
diw i = 1, 2, . . . , m, (12)
j =1

and the distance between the alternative i and the best condition Ab
v
u n
dib = t ∑ (tij − tbj )2 ,
u
i = 1, 2, . . . , m, (13)
j =1

where diw and dib are L2 -norm distances from the target alternative i to the worst and best
conditions, respectively.
• Step 6: Calculate the similarity to the worst condition:

diw
siw = , 0 ≤ siw ≤ 1, i = 1, 2, . . . , m. (14)
diw + dib

siw = 1 if and only if the alternative solution has the best condition; and siw = 0 if and only if the
alternative solution has the worst condition.
• Step 7: Rank the alternatives according to siw (i = 1, 2, . . . , m).

An open-source Python program [24] has been utilized for the TOPSIS analysis. Our input matrix
consists of 7 parameters, namely Precision(Equation (4)), Recall(Equation (5)) and mAP50-95(Equation
(7)) for each of the two models (Vehicle and License plate Detector) along with Total Inference Time in
milliseconds.

7.1.2. Criteria Weighting

Now the assignment of weights, Wj must be done on the basis of priority given to each metric
for our specific application. Here’s how the optimization of the combination of 16 different types of
system design must be carried out:

• Precision (Equation (4)): It becomes paramount for both car and license plate detection. False
positives, which can include detecting multiple cars, non-car objects, or non-plate objects, are
highly detrimental as they hinder accurate identification.
• mAP50-95 (Equation (7)): As IoU remains crucial for both car and license plate detection, mAP
at 95% IoU threshold has been considered as the one of the criteria in TOPSIS. Also, accurate
bounding box localization is essential for reliable downstream tasks like Optical Character
Recognition (OCR) applied to license plates.
• Recall (Equation (5)): Recall can be slightly relaxed for car detection compared to a multi-car
scenario. Missing a single car may be less frequent. However, a decent recall rate is still desirable.
Recall becomes less critical for license plate detection compared to car detection.

The weights Wi for both the models’ metrics are as follows:

• Precision: 20
• mAP50-95: 15
• Recall: 10

Weight Wi , of 1 has been assigned to Total Inference Time, in order to balance the trade-off
between speed and accuracy.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

21 of 25

7.1.3. Results
In Table 2, according to TOPSIS analysis, top 2 model combinations(highlighted in bold) are:

• TOPSIS Rank 1: YOLOv8s (Small) for Vehicle detection + YOLOv8m (Medium) for License Plate
detection
• TOPSIS Rank 2: YOLOv8m (Medium) for Vehicle detection + YOLOv8m (Medium) for License
Plate Detection.

Next, light intensity analysis will be performed for the top-ranked models according to TOPSIS
analysis.s

Table 2. Performance Comparison of Car Detector and License Plate Detector

VEHICLE DETECTION LICENSE PLATE DETECTION TOTAL INFERENCE TIME
S.NO. FPS TOPSIS
Model Precision Recall mAP50-95 Model Precision Recall mAP50-95 PER FRAME (Milliseconds) Rank

1 NANO 0.847 0.84 0.815 NANO 0.949 0.848 0.719 3.8 263.15 15
2 SMALL 0.942 0.853 0.747 6.3 158.73 7
3 MEDIUM 0.942 0.86 0.768 12.2 81.96 5
4 LARGE 0.946 0.859 0.787 19.2 52.08 8
5 SMALL 0.863 0.843 0.821 NANO 0.949 0.848 0.719 6.3 158.73 11
6 SMALL 0.942 0.853 0.747 8.8 113.63 3
7 MEDIUM 0.942 0.86 0.768 14.7 68.02 1
8 LARGE 0.946 0.859 0.787 21.7 46.08 6
9 MEDIUM 0.864 0.849 0.853 NANO 0.949 0.848 0.719 12.2 81.96 14
10 SMALL 0.942 0.853 0.747 14.7 68.02 4
11 MEDIUM 0.942 0.86 0.768 20.6 48.54 2
12 LARGE 0.946 0.859 0.787 27.6 36.23 9
13 LARGE 0.874 0.854 0.849 NANO 0.949 0.848 0.719 19.2 52.08 16
14 SMALL 0.942 0.853 0.747 21.7 46.08 12
15 MEDIUM 0.942 0.86 0.768 27.6 36.23 10
16 LARGE 0.946 0.859 0.787 34.6 28.9 13

7.2. Light Intensity Analysis:

For simulating the varying lighting conditions, a controllable light source along with an android
application [25] has been used to measure the light intensity at the region of interest. A smartphone
camera, coupled with Camo Studio [26], has been utilized to ensure high-quality footage and
integration with OpenCV library. The camera specifications [27] are as follows:

• Main Camera: 12-megapixel (12MP)

• Aperture: f/1.8
• Lens Configuration: 5P (5-element lens)
• Autofocus Technology: Phase Detection Autofocus (PDAF)
• Video Output: 1080p at 30FPS with Electric image stabilization (EIS)

A model car and the Indian High Security Registration Plate (HSRP) of 1:18 scale have been
utilized to ensure size continuity in the simulation environment.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

22 of 25

Figure 12. Variation of light intensity.

According to Indian Standards [28], minimum level of luminance on the roads must be 30 lux, in
order to ensure safety of the commuters. Thats’s why the ambient lighting level has been reduced to
20 lux in the analysis shown in Table 3. This proves an estimate of how well the system perform in low
lighting conditions.

Table 3. Light Intensity versus Class Confidence Scores of Vehicle and License Plate for Top 2 Models
in TOPSIS analysis.

TOPSIS Rank 1 Combination TOPSIS Rank 2 Combination

Light Intensity Class Confidence Class Confidence Class Confidence Class Confidence
in lux Score for Vehicle Score for License Plate Score for Vehicle Score for License Plate
900 0.90 0.62 0.93 0.51
800 0.89 0.65 0.94 0.62
700 0.90 0.67 0.94 0.61
600 0.90 0.70 0.93 0.60
500 0.90 0.70 0.92 0.64
400 0.89 0.70 0.93 0.63
300 0.89 0.72 0.92 0.63
200 0.88 0.66 0.89 0.61
100 0.91 0.64 0.90 0.40
50 0.88 0.65 0.89 0.50
20 0.87 0.65 0.85 0.67

8. Discussion
This research investigates a ensembled YOLOv8 approach for vehicle detection and License Plate
Detection within smart parking applications. The approach aims to support distributed parking across
the streets (Figure 1) and organize the chaotic encroachment due to vehicles on the shoulders of roads.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

23 of 25

8.1. Advantages
The proposed approach offers significant advancements over existing methods:

• Enhanced Accuracy and Efficiency: The ensembled strategy focuses the computational resource
on the regions of interest for License Plate Detection. Comparing with full-frame processing, the
proposed strategy results in better accuracy and shorter inference time. The advantage of the
present system is notable because most embedded systems have limited computational resources,
making the proposed system more applicable to the smart parking application.
• Elevated License Plate Detection Precision: The integration of pre-processing techniques
specifically tailored for license plate images significantly improves License Plate Detection
accuracy. These techniques address lighting variations and enhance image resolution before
feeding data to the Tesseract OCR engine, ensuring robust character recognition.
• Data-Driven Model Selection with TOPSIS: The present study employed the TOPSIS technique
from MCDM for the optimal selection of the YOLOv8 combinations. The technique is flexible and
can calculate the final score of ensembled YOLOv8 system after considering the mAP(Equation
(7)), Precision(Equation (4)), Recall(Equation (5)), and inference speed. Stakeholders can select
the model that satisfies the specific priorities of their smart parking application.
• Robustness Under Varying Lighting Conditions: The evaluation of the top YOLOv8
combinations from the previous study against different lighting intensities is practical information
because stakeholders are interested in the robustness of the model in real life.
• Investigation of the YOLOv8 combination trade-off: The current research is expected to
provide useful information about the trade-off of having a model with various sizes along
with performance related to its computational complexity, accuracy, and inference speed. To the
best of the author’s knowledge, the optimal selection of the YOLOv8 model has not been the
subject of investigation, but the information can help stakeholders to find out which YOLOv8
residue type they should put in their smart parking implementation.

8.2. Limitations
However, for successful real-world deployment, certain limitations require further consideration:

• Computational Bottleneck: The ensembled approach inherently introduces additional processing

overhead compared to single-stage object detection methods.
• Pre-processing Sensitivity: The reliance on pre-processing methods for license plate readability
adds complexity to the system. Investigating alternative pre-processing approaches that are
more resilient to lighting variations or dynamically adjust parameters based on real-time image
characteristics warrants further exploration.
• Considerations for Model Selection with TOPSIS: The TOPSIS methodology relies on the
assignment of subjective weights to performance metrics. Justifying these weights and assessing
the sensitivity of TOPSIS to different configurations is essential for robust model selection.
Expanding the evaluation to encompass a broader range of lighting conditions encountered in
real-world parking scenarios would further enhance the system’s generalizability.
• Scope of Lighting Variation Evaluation: The current evaluation of lighting conditions might
not fully capture the entire spectrum of real-world scenarios. Including very low light or
direct sunlight scenarios would provide a more comprehensive understanding of the system’s
robustness across diverse lighting environments.
• Data Availability of Low congestion streets: Nearly accurate data of least congested streets or
least traffic prone areas should be available. In order to implement this approach of smart parking
system.
• Availability of Surveillance Device: Surveillance device should be available coupled with
suitable hardware in order to ensure desired functioning of this system.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

24 of 25

By acknowledging these limitations through alternative system designs, refined pre-processing

techniques, and broader evaluation criteria, future iterations of this research can further strengthen the
understanding of trade-offs inherent in designing efficient and reliable smart parking systems.

9. Conclusion and Future Scope

This research has laid the foundation for the next-generation of smart parking solutions which
are more accurate, efficient, and robust than their predecessors. The ensembled YOLOv8 approach
with targeted processing, the pre-processing techniques for OCR enhancement, and TOPSIS-driven
model selection are considerable advancements to the state-of-the-art of this field. Also the approach
of parking distributed across the streets and wastelands is expected to reduce traffic congestion and
enhancement of the better utilization of limited spaces. It provides an opportunity to organize the
unorganized parking and revenue collection for the administrative authorities of the cities. Specific
limitations in the proposed design, such as the potential computational bottlenecks of the two-stage
implementation, may be addressed through the use of more efficient models for detection, such as
YOLOv9 [29]. Moreover, regional fonts can be used to train specialized OCR engines, rather than
enhance them prior to training. As an addition google maps can be integrated for identification of
availability of parking space. These technologies can be integrated to mean other system designs
as well, and additional research may be required to identify the possible best options. The system
also needs to be generalized and tested in a wider variety of parking environments through refining
pre-processing techniques and evaluations to ensure that these can be applied to any real-world
parking scenario.

References
1. McCoy K (2023) Drivers spend an average of 17 hours a year searching for parking spots. 2017
2. Cookson G, Pishue B (2017) Searching for parking costs americans $73 billion a year
3. Jocher G, Chaurasia A, Qiu J (2023) Ultralytics yolov8. https://github.com/ultralytics/ultralytics
4. Wojke N, Bewley A, Paulus D (2017) Simple online and realtime tracking with a deep association
metric. In: 2017 IEEE International Conference on Image Processing (ICIP), IEEE, pp 3645–3649,
10.1109/ICIP.2017.8296962
5. Wojke N, Bewley A (2018) Deep cosine metric learning for person re-identification. In: 2018 IEEE Winter
Conference on Applications of Computer Vision (WACV), IEEE, pp 748–756, 10.1109/WACV.2018.00087
6. Cao J, Pang J, Weng X, et al (2023) Observation-centric sort: Rethinking sort for robust multi-object tracking.
In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 9686–9696
7. Sharma N, Baral S, Paing MP, et al (2023) Parking time violation tracking using yolov8 and tracking
algorithms. Sensors 23(13):5843
8. Ultralytics (2022) ultralytics/yolov5: v7.0 - YOLOv5 SOTA Realtime Instance Segmentation. https://github.
com/ultralytics/yolov5.com, 10.5281/zenodo.7347926, https://doi.org/10.5281/zenodo.7347926, accessed:
7th May, 2023
9. Nguyen DL, Vo XT, Priadana A, et al (2023) Car detection for smart parking systems based on improved
yolov5. Vietnam Journal of Computer Science 1:15
10. Hasan Yusuf F, A Mangoud M (2024) Real-time car parking detection with deep learning in different lighting
scenarios. International Journal of Computing and Digital Systems 15(1):1–9
11. Das S (2019) A novel parking management system, for smart cities, to save fuel, time, and money. In: 2019
IEEE 9th annual computing and communication workshop and conference (CCWC), IEEE, pp 0950–0954
12. Redmon J, Divvala S, Girshick R, et al (2016) You only look once: Unified, real-time object detection. In:
Proceedings of the IEEE conference on computer vision and pattern recognition, pp 779–788
13. Badla S (2014) Improving the efficiency of tesseract ocr engine
14. n (2023) cars dataset. https://universe.roboflow.com/n-b8dyr/cars-ydptb, https://universe.roboflow.com/
n-b8dyr/cars-ydptb, visited on 2024-03-30
15. projects M (2024) license plates dataset. https://universe.roboflow.com/merge-projects/license-plates-
igyvb, https://universe.roboflow.com/merge-projects/license-plates-igyvb, visited on 2024-03-30
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 18 April 2024 doi:10.20944/preprints202404.1216.v1

25 of 25

16. Smith R (2007) An overview of the tesseract ocr engine. In: Ninth international conference on document
analysis and recognition (ICDAR 2007), IEEE, pp 629–633
17. OpenCV (2015) Open source computer vision library
18. Clark A (2015) Pillow (pil fork) documentation. https://buildmedia.readthedocs.org/media/pdf/pillow/
latest/pillow.pdf
19. Ren S, He K, Girshick R, et al (2016) Object detection networks on convolutional feature maps. IEEE
transactions on pattern analysis and machine intelligence 39(7):1476–1481
20. Sawant M (2022) Pytesseract: Python-tesseract is an optical character recognition (ocr) tool for python. that
is, it will recognize and "read" the text embedded in images. https://pypi.org/project/pytesseract/
21. Kim J, Cho J (2021) Rgdinet: Efficient onboard object detection with faster r-cnn for air-to-ground surveillance.
Sensors 21(5):1677
22. Hwang CL, Yoon K (2012) Multiple attribute decision making: methods and applications a state-of-the-art
survey, vol 186. Springer Science & Business Media
23. Zavadskas EK, Zakarevicius A, Antucheviciene J (2006) Evaluation of ranking accuracy in multi-criteria
decisions. Informatica 17(4):601–618
24. Wikipedia (2024) TOPSIS — Wikipedia, the free encyclopedia. http://en.wikipedia.org/w/index.php?title=
TOPSIS&oldid=1185091184, [Online; accessed 02-April-2024]
25. Phuongpn (2023) Illuminance - lux light meter. https://phuongpndev.web.app/, [Accessed 02-04-2024]
26. Ltd. R (2020) Camo studio. https://reincubate.com/camo/, [Accessed 02-04-2024]
27. (2019) Realme 5 Smartphone. Realme Mobile Telecommunications (India) Private Limited
28. Institution IS (1981) Indian Standard CODE OF PRACTICE FOR LIGHTING OF PUBLIC
THOROUGHFARES. Indian Standards Institution, New Delhi, iS 1944-5 (1981)
29. Wang CY, Liao HYM (2024) YOLOv9: Learning what you want to learn using programmable gradient
information

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those
of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s)
disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or
products referred to in the content.