Received: October 18, 2020. Revised: December 20, 2020.

Advancements in Object Detection and Distance Measurement for Autonomous

Vehicles: A Comparative Analysis of YOLO

Muhammad Abdillah Rahmat1,4 Indrabayu2* Andani Achmad3

A. Ejah Umraeni Salam3

1
Doctoral Student of Electrical Engineering Hasanuddin University, Makassar, Indonesia
2
Department of Informatics Hasanuddin University Makassar, Indonesia
3
Department of Electrical Engineering Hasanuddin University Makassar, Indonesia
4
Informatics Almarisah Madani University Makassar, Indonesia
* Corresponding author’s Email: [email protected]

Abstract: Autonomous vehicles, which are developed to operate autonomously with minimal or no human
intervention, have the potential to improve efficiency and safety on roads greatly. This project aims to
customize object detection and distance measuring technologies in Makassar by utilizing the advancements in the
YOLO architecture, namely from version 5 to 8. By utilizing a dataset consisting of 1296 vehicle images, this study
has effectively enhanced the accuracy of vehicle detection, a critical factor in achieving precise distance
measurement. YOLOv5 has exceptional accuracy, achieving a precision of 96.30% and a mAP (mean Average
Precision) of 97.50%. On the other hand, YOLOv8 maintains consistency with a mAP of 94.80% across a wider
range of IoU (Intersection over Union) thresholds. The device has exceptional precision in measuring distances,
especially at close ranges, with a difference of merely 0.23 meters from the true distance. The findings
emphasize the significance of developing technology specifically designed for the local environment in order to
promote the progress of self-driving vehicles. This will contribute to academic research and enhance transportation
safety in Makassar.

Keywords: Deep Learning, Vehicle Detection, Estimation Detection, YOLO

1. Introduction and accuracy in object detection, enabling rapid data

analysis [1].
The introduction of autonomous vehicle This study will utilize the most recent versions
technology is a significant innovation of the current of the YOLO architecture, specifically YOLOv5,
century, serving as a crucial catalyst in transforming YOLOv6, YOLOv7, and YOLOv8, to detect
the transportation sector. This technology holds vehicles. These iterations of YOLO epitomize the
substantial promise in enhancing transportation's forefront of object-detecting technology,
effectiveness, security, and convenience, showcasing enhanced precision, velocity, and
particularly in addressing prevalent issues like effectiveness in comparison to its forerunners. Using
traffic congestion and accidents. Object detection these sophisticated algorithms, our objective is to
and distance measuring are crucial elements in the get the utmost precision in detecting vehicle objects,
autonomous vehicle system, as they play a vital role a crucial stage in the vehicle distance measurement
in allowing cars to understand and interact with their process.
surroundings. The YOLO (You Only Look Once) The dataset plays a vital role in the advancement
architecture has demonstrated exceptional efficiency of object recognition systems. In Fig. 1 As an
illustration, we have employed a dataset including
International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 2

1296 images that have been labelled with the label Makassar, situated in South Sulawesi, presents a
"vehicle". Figure 1 displays an image of the data diverse array of road conditions that require the
utilized in this research. This comprehensive dataset application of sophisticated technologies for
offers a reliable depiction of several traffic scenarios adaptation and modification. The attributes
that the system might encounter. Put simply, the encompass deteriorated road conditions, potholes,
dataset includes several scenarios that can help the accumulation of water during precipitation, and
system gain a deeper understanding and improved traffic congestion. Makassar, known for its lively
recognition of varied traffic conditions. culture and diverse socio-economic makeup, has a
unique traffic pattern characterized by a variety of
vehicles such as private cars, public transportation,
and motorcycles. The autonomous driving system
must surmount the increased intricacy of object
recognition and distance estimation.
Figure 1. Dataset Furthermore, the implementation of this
technology is greatly impacted by cultural and
In addition, we have performed 100 epochs socioeconomic factors. The cultural inequalities
during the model training phase. Ensuring that the and road user conduct in Makassar, Indonesia, have
model can learn the diverse range of conditions and a significant impact on the interactions between road
variants in the dataset is a crucial step. Through users. Considering this aspect is essential for
undergoing multiple epochs, the model can adjust to developing effective algorithms for object detection
the data and progressively enhance its performance. and distance measurement. Extreme variations in
This approach enhances the model's accuracy and weather conditions, such as heavy precipitation, can
dependability in identifying vehicle objects in lead to malfunctions in the sensors and cameras of
different traffic scenarios. self-driving vehicles, resulting in new problems that
By utilizing an extensive dataset and conducting need to be addressed.
a meticulous training process, we may enhance the The objective of this study is to conduct a
development of a superior and more precise system comparative analysis of the YOLO architecture,
for identifying vehicle objects. Undoubtedly, this focusing primarily on its performance in detecting
will greatly improve traffic safety and efficiency on objects in the traffic environment and under various
the roadways. road conditions in the city of Makassar. We will
To improve the precision of vehicle object evaluate the effectiveness and reliability of object
detection, preliminary processing measures are recognition and distance measurement technology in
implemented. The procedures involve the removal overcoming the specific challenges faced in this
of noise, standardizing lighting conditions, and urban environment. The project's objective is to
rectifying distortions induced by camera lenses. produce valuable insights and effective solutions to
Through the process of refining the data, the object improve the performance of autonomous vehicles in
detection algorithm may operate at its highest level Makassar. Furthermore, it aims to make a valuable
of efficiency, accurately detecting automobiles with contribution to the scholarly body of knowledge on
enhanced precision. Consequently, this enhances autonomous transportation technology.
the precision of vehicle distance estimation, as the To successfully implement autonomous vehicle
system depends on good object detection to technology in Indonesia, particularly in Makassar, it
precisely calculate the distance between vehicles is crucial to have a thorough comprehension of the
[2]. local environment. When designing and deploying
For distance measurement, we will employ a autonomous vehicle systems, it is crucial to consider
measurement technique that utilizes stereo cameras. the distinct traffic circumstances, road
Stereo cameras enable the system to acquire images infrastructure, and cultural factors of the region.
from two distinct perspectives, thereby supplying The YOLO architecture, known for its demonstrated
the essential depth data required for calculating the efficacy and accuracy in object detection, offers a
distance between objects in the image. Our strong basis for tackling the difficulties presented by
objective is to improve the precision of vehicle the traffic environment and road conditions in
distance measurement in the autonomous vehicle Makassar.
system by utilizing stereo cameras. To summarize, the use of autonomous vehicle
Implementing autonomous vehicle technology in technology signifies a substantial advancement in
Indonesia is difficult because of the country's unique the progression of the transportation sector. The
traffic conditions and road infrastructure. potential advantages of this technology, in terms of
International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 3

enhanced productivity, security, and convenience in (YOLO), in terms of speed and efficiency. The
transportation, are significant. Nevertheless, the COCO (Common Object in Context) dataset was
effective execution of autonomous vehicle systems utilized for the comparison analysis, wherein the
necessitates a meticulous evaluation of the specific performance of both algorithms was assessed in
circumstances, encompassing traffic conditions, terms of accuracy, precision, and speed. YOLO-v3
road infrastructure, and cultural elements. The outperforms SSD in same testing circumstances, as
objective of this study is to offer important insights indicated by the findings [5].
and practical ideas to improve the performance of Finally, Paweł Kowalczyk et al.'s research
self-driving vehicles in the city of Makassar. This centers on examining the decline in detection
research will contribute to the existing academic accuracy caused by simulated camera flaws, which
literature on autonomous transportation technology. occur during the evaluation of self-driving cars.
The researchers employ statistical techniques and
2. Related Work topological analysis to offer comprehensive insights
into the distribution of quality in this comparative
Autonomous vehicles have emerged as a
experiment [6].
dynamic field of study, encompassing various
For creating a vehicle identification and distance
disciplines including artificial intelligence (AI), the
measurement system tailored to the conditions in
Internet of Things (IoT), embedded systems, and
Makassar, Indonesia, additional refinement and
control engineering. An essential element of
enhancement of current algorithms and technologies
autonomous navigation is the capacity to identify
are required. The traffic conditions and
automobiles, pedestrians, or other obstructions on
infrastructure in Makassar have distinct
the road and ascertain their proximity to the
characteristics that may vary from circumstances in
autonomous vehicle. The main algorithms
Western countries or other prominent cities in
employed for this objective utilize cameras and
Indonesia. Hence, it is imperative to conduct
Light Detection and Ranging (LiDAR) data.
research and development on a vehicle detection and
Another classification of algorithms involves the
distance measuring system that is well-suited for the
amalgamation of these two sets of sensor data.
specific conditions of the area. This is essential to
Sensor fusion networks utilize 2D camera pictures
guarantee the precision and dependability of the
and LiDAR point clouds as input to generate 3D
system in its role of assisting autonomous vehicles
bounding boxes as the output of detection.
in the region.
The study conducted by Shafaq Sajjad and
colleagues assesses the efficacy of three object
3. Methodology
detection techniques, taking into consideration the
nature of the input data. The article presents a This paper presents a methodical and organized
comparative analysis of three object identification approach to examine progress in object
networks, considering variables such as accuracy, identification and distance measurement for self-
performance in obstructed situations, and computing driving vehicles. The main emphasis is on doing a
complexity. The YOLOv3, BEV network, and comparative analysis of different iterations of
Point Fusion models underwent training and testing YOLO (You Only Look Once). The main goal of
using the KITTI dataset, which serves as a this technique is to offer a comprehensive
benchmark. The results demonstrate that the sensor comprehension and analysis of the advancements in
fusion network outperforms the single-input object detection and distance measuring
networks [3]. technologies, as well as their practical use in
Moreover, the study conducted by Muhammed autonomous cars to improve navigation safety and
Muzammul and Xi Li offers a comprehensive efficiency. We have delineated the fundamental
examination of the difficulties and remedies in stages in our study methodology, commencing with
object detection using computer vision, employing data pre-processing and culminating in the
diverse methodologies. The researchers perform a assessment of the efficacy of the suggested
comparative investigation of various object framework and various iterations of YOLO.
identification methods, including R-CNN, Fast R- Through the utilization of this methodological
CNN, Faster R-CNN, YOLO, and SSD [4]. approach, our objective is to provide a significant
The study conducted by Peddinti Mounika et al. contribution to the existing body of knowledge in
aims to compare the performance of two prominent the field of autonomous cars and facilitate future
object identification algorithms, namely Single Shot study in this domain [7].
identification (SSD) and You Only Look Once

International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 4

3.1 Pre-processing will aid in the advancement of more precise and

Figure 2. Pre-processing
Data pre-processing is the initial stage in our dependable models for detecting objects and
methodology, and it plays a vital role in measuring distances.
guaranteeing the quality and uniformity of the data
utilized for training and testing our models. It can
be seen in Fig. 2. This stage encompasses multiple 3.2 Proposed Architecture
sub-processes: In this paper, we offer an architecture for
 Data Augmentation vehicle object detection and distance measurement
To increase the variety of our dataset and from autonomous cars, which consists of multiple
strengthen the resilience of our models, we carry out main stages. It can be seen in Fig. 3. They are as
data augmentation. This entails generating different follows:
iterations of the initial data by employing methods  Dataset Collection Using Stereo Cameras
such as rotation, mirroring, and resizing [8]. The first phase entails collecting datasets
 Normalization and Contrast Enhancement through the utilization of stereo cameras. Stereo
The data is subsequently normalized to cameras enable precise depth acquisition, a critical
standardize all pixel values within a defined range, factor in determining the distance of objects from
often ranging from 0 to 1. Ensuring that our models autonomous cars. The dataset comprises a diverse
are not biased towards specific values and can collection of photos and videos depicting different
successfully learn from the data is crucial. We road scenarios, automobiles, and pedestrians,
augment the contrast of the photos to enhance the captured under varying lighting and weather
discernibility of the items. This stage is crucial in circumstances.
situations where the lighting conditions are  Pre-processing
suboptimal, and the objects are not distinctly Following the collection of the dataset, the
apparent [9]. subsequent stage involves data pre-processing.
 Noise Reduction This process entails employing data augmentation
The data is subjected to several filtering techniques to enhance the diversity of the dataset,
algorithms to minimize any noise present. This normalizing pixel values to a specific range,
phase is essential for enhancing the quality of the enhancing contrast to improve object
photos and increasing the recognizability of the distinguishability, reducing noise to enhance image
items [10]. clarity, segmenting objects to separate them from
 Object Segmentation the background, and enhancing resolution to ensure
The photos undergo object segmentation to clear visibility of objects, even at a distance.
distinguish them from the backdrop. Performing  Feature Extraction
this step is crucial in directing our models' attention After the data has received processing, the
towards the specific objects of interest, hence subsequent stage involves extracting features.
enhancing the precision of object detection [11]. Currently, we employ deep learning methods to
 Resolution Enhancement extract crucial characteristics from photos, which
The images resolution is heightened to will be utilized for object detection and distance
guarantee optimal visibility of objects, even when calculation. This method utilizes convolutional
they are situated at a considerable distance. This layers to detect patterns and attributes of objects in
step is essential for precise distance measuring [12]. photos, together with pooling layers to decrease data
Once the data has been pre-processed, it is dimensionality and accelerate the learning
subsequently inputted into our deep learning models procedure.
that utilize the YOLO (You Only Look Once)  Classification
architecture for both object detection and distance Once the features have been recovered, the
measurement. Our objective is to provide a high- subsequent stage involves categorization.
quality dataset that accurately reflects the real-world Currently, we employ deep learning models to
situations encountered by autonomous vehicles, by categorize items depicted in photos as either cars or
adhering to these sub-processes. Consequently, this non-vehicles. The models are trained by utilizing

International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 5

the processed dataset and extracted features. This is swift and precise object detection, such as
accomplished through techniques like transfer autonomous vehicles and real-time video analysis.
learning, which allows for the utilization of The structure of YOLO comprises a
knowledge acquired from earlier tasks. convolutional neural network (CNN) that extracts
 Vehicle Object Detection characteristics from the input image. The network

Figure 3. The research process

After the conclusion of classification, the partitions the image into a grid, where each grid cell
subsequent stage involves the detection of vehicle assumes the responsibility of predicting items
objects. Currently, we employ the YOLO (You contained within its respective bounds. Every
Only Look Once) model to detect and categorize individual cell makes predictions for numerous
automobile entities in photos. YOLO is an efficient bounding boxes and their corresponding probability
object detection technique that can rapidly recognize for different classes. The bounding boxes are
and categorize things, making it well-suited for defined by their central coordinates, width, height,
autonomous vehicle systems that demand swift and a confidence score that indicates the precision of
reactions. The YOLO model is trained using the the prediction. The class probabilities indicate the
pre-processed dataset, extracted features, and probability of the discovered object being a member
categorized data. of a specific class.
 Vehicle Distance Measurement
Once the vehicle objects have been identified, The general prediction formula for YOLO can be
the last stage involves determining the precise described as:
distance between the cars and the autonomous
P ( Classi|Object ) × P ( Object ) × IOU pred (1)
truth
vehicle. Currently, we utilize the depth data
acquired from the stereo cameras and measuring
techniques to accurately determine the distance Where:
between the autonomous vehicle and the observed  P ( Classi|Object ) is the probability that the
vehicle objects. The autonomous car utilizes this
distance data to make precise navigation choices, object is a member of class i.
 P ( Object ) is the probability of an object being
including collision avoidance and keeping a secure
separation from other vehicles. present within the bounding box.
truth
 IOU pred are the Intersection over Union (IoU)
3.2.1. YOLO measures the overlap between the predicted
YOLO (You Only Look Once) is an advanced bounding box and the ground truth bounding
object detection system that use a single neural box.
network to directly forecast bounding boxes and
class probabilities from whole images in a single The YOLO framework has seen multiple
assessment. This methodology exhibits a notable iterations, each subsequent edition contributing
divergence from conventional methods, which enhancements in terms of precision, velocity, and
usually entail distinct elements for region suggestion usefulness. The latest iterations, YOLOv5,
and classification. The unified model of YOLO has YOLOv6, YOLOv7, and YOLOv8, have included
several benefits, including a streamlined system several upgrades, including enhanced backbone
capable of real-time image processing. This makes it designs, superior feature extraction, and more
especially well-suited for applications that demand efficient bounding box regression approaches.

International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 6

These technological breakthroughs have led to the Through the utilization of this technique, we can
development of models that exhibit not only accurately determine the distance between the
increased speed but also enhanced accuracy in vehicles. This precise measurement may then be
recognizing objects across various scales and used by the autonomous vehicle to make well-
settings. informed decisions on navigation.
YOLO is an innovations development in object 3.3 Evaluation Metrics
detection, providing a comprehensive and In addition to thoroughly evaluate the
immediate solution that can be applied to various effectiveness of our suggested framework, we utilize
situations. The architecture of this network a collection of evaluation measures that are widely
integrates feature extraction, bounding box accepted and commonly used in the domain of
prediction, and class probability estimation into a object identification and distance measuring. These
single system, establishing a new benchmark for metrics are essential for measuring the precision,
efficiency and efficacy in object detection tasks. dependability, and overall efficiency of the system.
The ongoing enhancements and revisions to the The evaluation metrics employed in this study are:
YOLO architecture provide additional evidence of  Precision
its significance and promise for future utilization in Precision is an essential measure that quantifies
other fields. the level of accuracy in detecting objects. The term
refers to the proportion of correctly identified
3.2.2. Measurement Method objects (true positive detections) divided by the sum
The distance measurement method is a of both correctly and wrongly identified objects
fundamental component of object detecting systems (true positive and false positive detections,
used in autonomous cars. Accurate distance respectively). Mathematically, precision is
measuring is essential to guarantee the vehicle can determined using the following formula:
make accurate navigation judgments, such as
avoiding crashes and maintaining a secure distance True Positive
Precision= (3)
from other objects. True Positives+ False Positives
This study implements a stereo triangulation-
based distance measurement method, which A higher precision rating indicates a system's greater
involves the use of two stereo cameras to determine accuracy in recognizing objects, while minimizing
the distance between an item and the vehicle. This false positives. Accuracy is especially crucial in
technique operates by detecting matching points situations where the consequences of incorrect
between two images captured by two distinct positive results are expensive.
cameras. By utilizing these matching points, we
 Recall
can determine the disparity, which is subsequently
Recall, generally referred to as sensitivity or true
employed to measure the distance between the item
positive rate, is a quantitative measure that assesses
and the vehicle.
the system's capacity to accurately identify all
The formula for determining the distance of an
pertinent objects present in the scene. The
object using the stereo triangulation technique can
calculation involves dividing the number of
be expressed as:
correctly identified items (true positives) by the sum
f ×B (2) of the true positives and the objects that were missed
Distance=
d (false negatives). The formula for recall is as
follows:
Where:
 Distance is to the measurement of the object's True Positive (4)
position relative to the vehicle.
Recall=
True Positives+ False Negatives
 The variable f represents the focal length of the
camera, measured in pixels. A greater recall score indicates that the system is
 B is the fundamental distance between the two proficient in capturing all pertinent objects, while
stereo cameras, measured in millimeters. minimizing the occurrence of false negatives.
 The disparity, denoted as d, is the horizontal Recall is of utmost importance in situations where
difference in coordinates of comparable spots the failure to detect an object might result in major
between the two images, measured in pixels. repercussions.
 mAP 50

International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 7

Mean Average Precision at 50% Intersection lasted for 100 epochs. We utilized an NVIDIA Tesla
over Union (mAP 50) is a comprehensive metric T4 GPU for data processing and model training.
used to assess the precision of object detection. An essential stage in this experiment involved
The calculation involves determining the average the creation of pre-processing. Pre-processing is
precision using a 50% Intersection over Union (IoU) the method of preparing and purifying data before
criterion. IoU is a metric that calculates the degree its utilization in model training. The objective of
of overlap between the predicted bounding box and pre-processing is to enhance the data's quality,
the actual bounding box. The formula for mean hence facilitating the model's learning process and
Average Precision at 50 (mAP 50) is as follows: generating more precise predictions.
We have generated a table that compares
N various models for object detection based on the
1 (5)
mAP 50=
N
∑ Precisioni data received from the vehicle object detection.
i=1
Four detector models employed were YOLOv5,
N represents the total count of object classes. A
YOLOv6, YOLOv7, and YOLOv8. Every detector
higher mean Average Precision (mAP) at 50%
model demonstrated distinct precision, recall, and
overlap implies that the system has a greater level of
mAP values. The pre-processing development
accuracy in detecting objects, with a substantial
phase revealed that the YOLOv5 model had a high
overlap between the predicted bounding boxes and
level of precision, measuring at 95.90%, and a
the ground truth bounding boxes.
commendable recall value of 86%. The YOLOv7
 mAP 50-95 model achieved the highest recall rate, reaching
Mean Average Precision at 50-95% Intersection 94%. Table 1 provides more comprehensive results.
over Union (mAP 50-95) is a statistic that builds Following the pre-processing of the
upon the mAP 50 metric. It assesses the precision development phase, there was an improvement in
of object detection at different IoU levels, spanning the precision, recall, and mAP values for each
from 50% to 95%. This statistic offers a more detector model. The YOLOv5 model achieved the
thorough assessment of the system's performance highest precision and recall rates, specifically
across various degrees of overlap with the ground 96.30% and 92.20% respectively. The YOLOv6
truth. The formula for calculating the mean average model achieved the highest recall rate, reaching
precision (mAP) for a range of 50-95 is as follows: 98%. Table 2 provides more comprehensive
N K
results.
1
mAP 50−95= ∑1
N i=1 K
∑ Precisioni , j (6) During the evaluation of the distance detecting
system for motorbikes prior to pre-processing
J=1 ❑
development, the system accurately estimated a
N represents the total count of object classes, distance of 4.68 meters when the real distance was 5
while K represents the total count of IoU thresholds. meters. The system's prediction deviated by 0.47
A higher mean average precision (mAP) score meters from the actual distance of 10 meters. The
between 50-95 shows that the system has a greater system's prediction of 17.47 meters deviated from
level of accuracy in detecting objects, considering the actual distance of 15 meters. Nevertheless,
various amounts of overlap with the ground truth when positioned at a precise distance of 20 meters,
bounding boxes. the system was unable to perceive the distance.
These metrics collectively offer a comprehensive This suggests that the system's ability to detect
assessment of the system's ability to recognize distances at longer ranges is still limited. Table 3
vehicle objects and accurately measure their provides more comprehensive results.
distance from the autonomous vehicle. Through the During the evaluation of the car distance
analysis of these indicators, we may identify specific detecting system prior to pre-processing
areas that need improvement and then optimize the development, the system accurately predicted a
system to better its overall performance. distance of 4.94 meters when the actual distance was
5 meters. At a measured distance of 10 meters, the
4. Result and Experiment system estimated a value of 10.23 meters. The
system's prediction of 17.22 meters deviated from
This study involved conducting trials to improve the actual distance of 15 meters. However, when
the precision of distance detection systems in the system was positioned at a precise distance of 20
motorcycles and cars, as well as object detection meters, it was unable to accurately perceive the
systems. The trials employed a dataset consisting distance. This suggests that the system's ability to
of 1296 images, and the model training process detect distances at longer ranges is still limited.
International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 8

Table 3 provides a more comprehensive display of distance of 5 meters. The system's prediction of the
the results. distance was 9.89 meters, which was somewhat less
Following the pre-processing of development, than the actual distance of 10 meters. The system's
the evaluation of the distance sensing system for prediction of the distance was 15.23 meters, which
motorbikes exhibited notable enhancement. The closely matched the actual distance of 15 meters.
method accurately predicted a distance of 4.86 However, even when positioned at a precise distance
meters, which was only slightly less than the actual

Table 1. Experimental vehicle detection results before to pre-processing optimization.

Detector Backbone Precision Recall mAP 50 mAP 50-95
YOLOv5 CSPDarknet53 95,90% 86% 92,90% 90,10%
YOLOv6 EfficientRep 87% 87% 90,20% 89,20%
YOLOv7 ELAN+CBS 84,80% 94% 92,10% 91,40%
YOLOv8 Resnet50 90,40% 88% 92,70% 92,30%

Table 2. Experimental vehicle detection results after to pre-processing optimization.

Detector Backbone Precision Recall mAP 50 mAP 50-95
YOLOv5 CSPDarknet53 96,30% 92,20% 97,50% 93,80%
YOLOv6 EfficientRep 90% 98% 91,20% 90,40%
YOLOv7 ELAN+CBS 92,20% 88,70% 93,40% 92,40%
YOLOv8 Resnet50 94,50% 91,70% 95,50% 94,80%

of 20 meters, the device remained unable to obstacles to be surmounted, particularly in

determine the distance. Table 4 provides a more enhancing the system's precision in identifying
comprehensive overview of the data. distances at extended ranges.
Similarly, during the evaluation of the car The advancement of pre-processing techniques
distance detecting system after pre-processing in vehicle object detection has greatly impacted the
development, it was observed that when the actual accuracy of distance detection, specifically in
distance was 5 meters, the system accurately relation to vehicle attributes in Indonesia,
predicted a distance of 4.97 meters. The system's particularly in Makassar. In Indonesia, especially
prediction for the actual distance of 10 meters was in major urban areas such as Makassar, there is a
10.12 meters. The system's prediction of the wide range of vehicles in operation, including
distance was 15.42 meters, which was only 0.42 motorbikes, private cars, and public transportation
meters more than the actual distance of 15 meters. such as minibusses and buses. Vehicles in Indonesia
Nevertheless, even when positioned at a precise typically exhibit smaller dimensions and lower
distance of 20 meters, the device remained incapable weight in comparison to their counterparts in
of detecting the distance. Table 4 provides a more Western nations. Furthermore, traffic flow in
detailed display of the results. The study shows Indonesia is characterized by higher density and
instances of vehicle detection and measurement of greater irregularity, as a significant number of cars
vehicle distances, as shown in Fig 2. are present on the roadways.
The research has demonstrated that the
implementation of pre-processing techniques in
vehicle object detection has greatly improved the
accuracy of distance detection. Through the
implementation of pre-processing techniques, the
picture data of cars utilized for object detection
underwent refinement, resulting in improved clarity
Figure 4. Example of vehicle detection and vehicle
and enhanced ease of processing for the model.
distance
This led to improved and more precise distance
Based on these assessments, it can be inferred
detection, which is vital considering the specific
that the implementation of pre-processing
features of the vehicle and traffic patterns in
techniques has resulted in a favorable enhancement
Indonesia.
in the precision of distance sensing systems for
bikes and cars. Nevertheless, there remain still
International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020. 9

Moreover, the distance detection system, which Therefore, the study's novelty is in the
incorporates pre-processing techniques, can be advancement of pre-processing techniques for
suggested for real-time implementation prior to the vehicle object detection. This contribution greatly
stage of determining the relative velocity of enhances the research and development of future
vehicles. This technique enables us to correctly and technologies for detecting vehicle distance speed,
rapidly identify the distance of vehicles, hence
facilitating more effective decision-making during
driving. Undoubtedly, this will enhance driving
safety and mitigate the incidence of road accidents,
particularly in big urban areas such as Makassar.

Table 4. Evaluation of Distance Detection Systems for Motorcycles and Cars after pre-
processing evaluated
Table 3. Evaluation of Distance Detection Systems for Motorcycles and Cars before pre-
Motorcycle Distance Detection (Meters)
processingCar Distance Detection (Meters)
evaluated
Actual Predictions
Motorcycle Distance Detection (Meters) Actual Predictions
Car Distance Detection (Meters)
5Actual 4,86
Predictions 5
Actual 4,97
Predictions
10
5 9,89
4,68 10
5 10,12
4,94
15
10 15,23
9,53 15
10 15,42
10,23
20
15 Not Detected
17,47 20
15 Not Detected
17,22
20 Not Detected 20 Not Detected
particularly in the specific context of vehicle advancements were achieved through the
characteristics and traffic patterns in Indonesia. utilization of a comprehensive dataset and
The examination of the originality of pre- meticulous training techniques.
processing advancement in vehicle object detection The pre-processing phases of data, which
suggests that this stage is a vital milestone in include augmentation, normalization, contrast
enhancing the precision of distance detection enhancement, noise reduction, object segmentation,
systems. and resolution improvement, were found to be
Pre-processing enhances the quality of vehicle essential in improving the quality of data required
image data by removing unwanted disturbances and for deep learning. The architecture
improving its optimization. This, in turn, enables the shows outstanding proficiency in vehicle detection
model to acquire superior learning capabilities and with great accuracy and precise distance
generate more precise predictions. This is crucial measurement, by applying depth information from
due to the intricate and unpredictable vehicle stereo cameras.
attributes and traffic patterns in Indonesia in contrast YOLOv5 showed exceptional performance in
to Western nations. Hence, the advancement of vehicle detection, with a precision rate of 96.30%
pre-processing in vehicle object identification is a and a recall rate of 92.20%. Additionally, it attained
crucial stage that greatly influences the accuracy of a mean Average Precision (mAP) of 97.50% at an
distance detection. It is recommended to implement Intersection over Union (IoU) threshold of 50%. In
this phase in real-time before detecting the relative contrast, YOLOv8 demonstrated greater stability in
speed of vehicles. its performance across different testing situations,
achieving a mean average precision (mAP) of
5. Conclusion 94.80% across IoU thresholds ranging from 50% to
95%. The results suggest that YOLOv5
In this paper, shows the significance of
demonstrates a high level of accuracy in identifying
modifying object identification and distance
vehicles in specific situations, while YOLOv8
measurement technology for autonomous cars to suit
exhibits more consistent performance across a wide
the particular traffic conditions and infrastructure of
variety of testing conditions. This represents notable
Makassar City. An exhaustive assessment of the
progress in the creation of dependable detection
YOLO architectural series, ranging from YOLOv5
systems for self-driving cars.
to YOLOv8, uncovered substantial enhancements in
The proposed technology demonstrated
the precision of vehicle detection. These
exceptional accuracy in measuring distances.
International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020.
10

Before undergoing pre-processing, the system References

demonstrated a remarkable ability to accurately
anticipate distances, such as estimating a distance of [1] J. Estrada, G. Opina Jr, and A. Tripathi, “Object
4.94 meters for an actual distance of 5 meters. and Traffic Light Recognition Model
Following pre-processing, the system's accuracy Development Using Multi-GPU Architecture
notably enhanced, achieving a distance prediction of for Autonomous Bus,” in Frontiers in Artificial
15.23 meters for an actual distance of 15 meters, Intelligence and Applications, A. J. Tallón-
demonstrating a minimal variance of merely 0.23 Ballesteros, Ed., IOS Press, 2021. doi:
meters. The results validate the system's 10.3233/FAIA210286.
dependability in measuring shorter distances, while [2] A. Fregin, J. Muller, and K. Dietmayer, “Three
it still has difficulties when it comes to greater ways of using stereo vision for traffic light
distances. recognition,” in 2017 IEEE Intelligent Vehicles
This paper introduces significant advancements Symposium (IV), Los Angeles, CA, USA: IEEE,
necessary for the effective deployment of Jun. 2017, pp. 430–436. doi:
autonomous cars, including obstacles such as 10.1109/IVS.2017.7995756.
varying traffic patterns, socio-economic and cultural [3] S. Sajjad, A. Abdullah, M. Arif, M. U. Faisal,
influences, and unpredictable weather conditions. M. D. Ashraf, and S. Ahmad, “A Comparative
These findings possess the capacity to improve Analysis of Camera, LiDAR and Fusion Based
safety and effectiveness in Makassar's transportation Deep Neural Networks for Vehicle Detection,”
system, while also providing valuable insights to the IJIST, vol. 3, no. 5, pp. 177–186, Jan. 2022, doi:
academic literature in the field of autonomous 10.33411/IJIST/2021030514.
transportation technology. This highlights the [4] M. Muzammul and X. Li, “A Survey on Deep
significance of modifying technology to suit specific Domain Adaptation and Tiny Object Detection
local requirements and environments. Challenges, Techniques and Datasets,” ACM
Comput. Surv., p. 1122445.1122456, May 2022,
Conflicts of Interest (Mandatory) doi: 10.1145/1122445.1122456.
[5] P. Mounika and C. L. Narayana, “A
Declare conflicts of interest or state “The COMPARATIVE STUDY OF YOLO AND
authors declare no conflict of interest.” Authors SSD OBJECT DETECTION ALGORITHMS,”
must identify and declare any personal vol. 9, no. 11, 2021.
circumstances or interest that may be perceived as [6] P. Kowalczyk, P. Bugiel, M. Szelest, and J.
inappropriately influencing the representation or Izydorczyk, “Fault injection in optical path -
interpretation of reported research results. detection quality degradation analysis with
Wasserstein distance,” in 2021 25th
Author Contributions (Mandatory) International Conference on Methods and
For research articles with several authors, a short Models in Automation and Robotics (MMAR),
paragraph specifying their individual contributions Międzyzdroje, Poland: IEEE, Aug. 2021, pp.
must be provided. The following statements should 121–126. doi:
be used as follows: “Conceptualization, XXX and 10.1109/MMAR49549.2021.9528441.
YYY; methodology, XXX; software, XXX; [7] U. Goenka, A. Jagetia, P. Patil, A. Singh, T.
validation, XXX, YYY, and ZZZ; formal analysis, Sharma, and P. Saini, “Threat Detection in Self-
XXX; investigation, XXX; resources, XXX; data Driving Vehicles Using Computer Vision,” in
curation, XXX; writing—original draft preparation, Machine Learning, Image Processing, Network
XXX; writing—review and editing, XXX; Security and Data Sciences, vol. 946, R.
visualization, XXX; supervision, XXX; project Doriya, B. Soni, A. Shukla, and X.-Z. Gao,
administration, XXX; funding acquisition, YYY”, Eds., in Lecture Notes in Electrical Engineering,
etc. Authorship must be limited to those who have vol. 946. , Singapore: Springer Nature
contributed substantially to the work reported. Singapore, 2023, pp. 617–630. doi:
10.1007/978-981-19-5868-7_45.
Acknowledgments [8] S. Aleem, T. Kumar, S. Little, M. Bendechache,
R. Brennan, and K. McGuinness, “Random
Acknowledgments are to show that the article is Data Augmentation based Enhancement: A
supported by what organization. For example, “This Generalized Enhancement Approach for
work was supported by the National Nature Science Medical Datasets.” arXiv, Oct. 03, 2022.
Foundation under Grant No. 405”.
International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx
 Received: October 18, 2020. Revised: December 20, 2020.
11

Accessed: Nov. 04, 2023. [Online]. Available:

http://arxiv.org/abs/2210.00824
[9] R.V.R. & J.C. College of Engineering, G. Rao,
A. Srikrishna, R.V.R. & J.C. College of
Engineering, N. Vadaparthi, and Maharaj
Vijayaram Gajapathi Raj College of
Engineering, “Weber’s Law based Camera
Response Model for Low Light Image Contrast
Enhancement,” IJIES, vol. 12, no. 5, pp. 33–42,
Oct. 2019, doi: 10.22266/ijies2019.1031.04.
[10]I. Elawady and C. Özcan, “A new effective
denoising filter for high density impulse noise
reduction,” Turkish Journal of Electrical
Engineering and Computer Sciences, vol. 30,
no. 4, pp. 1388–1403, May 2022, doi:
10.55730/1300-0632.3855.
[11]Q. Guo, J. Liu, and M. Kaliuzhnyi, “YOLOX-
SAR: High-Precision Object Detection System
Based on Visible and Infrared Sensors for SAR
Remote Sensing,” IEEE Sensors J., vol. 22, no.
17, pp. 17243–17253, Sep. 2022, doi:
10.1109/JSEN.2022.3186889.
[12]M. Jiafa, H. Wei, and S. Weiguo, “Target
distance measurement method using monocular
vision,” IET image process, vol. 14, no. 13, pp.
3181–3187, Nov. 2020, doi: 10.1049/iet-
ipr.2019.1293.

International Journal of Intelligent Engineering and Systems, Vol.x, No.x, 20xx DOI: 10.22266/ijies2019.xxxx.xx