YOLO NAS-powered Intelligent Vehicle Tracking

and Monitoring System for Enhanced Parking

Management in Urban Environments
space management and use by utilizing the power of
YOLO NAS, a cutting-edge object identification algorithm
Abstract— The increasing complexity of urban known for its speed and accuracy.
infrastructure coupled with increasing vehicle ownership
This paper explores the technical details of our
has led to a growing challenge in efficiently managing
suggested solution, demonstrating how computer vision,
parking spaces in busy environments such as shopping
real-time monitoring, and AI-driven tracking come together
centers. A recurring problem drivers face is the to offer a reliable and effective way to locate and identify
difficulty in locating their vehicles parked in large parked cars. In addition to improving user experience per
parking lots. This article presents a novel solution to se, our system's deployment could improve parking space
address this problem by implementing an AI-based management in general and give facility managers useful
intelligent tracking and monitoring system. Our information for streamlining operations and optimizing
proposed system utilizes the state-of-the-art object layouts.We will go over the design, features, and
detection algorithm YOLO NAS (You Only Look Once advantages of our Smart AI-Based Vehicle Monitoring
NAS) to accurately identify and track vehicles in the System in the sections that follow. We want to emphasize
parking area. By integrating DeepSORT (Deep, Simple how this system may improve parking situations, lessen
Online Real-Time Tracking), the Python programming annoyances, and help create a more responsive and
language, and the real-time monitoring capabilities of intelligent urban infrastructure.
the Pytorch framework, our system helps users quickly Because of its exceptional speed and accuracy in object
locate their parked vehicles, eliminating frustration and detection, the YOLO NAS model is used to process video
wasted time in finding a parked one Automobile.With feeds from strategically placed surveillance cameras within
this innovative approach, we aim to alleviate the the parking lot in an efficient manner. The system gives
common problem of lost or forgotten parked cars, consumers the most recent position information on their
providing users with a seamless and stress-free parking parked automobiles by accurately recognizing and tracking
experience while contributing to more effective and individual vehicles using sophisticated computer vision
intelligent parking management in commercial techniques.Our intelligent AI-based car monitoring system
establishments. The results of our proposed solution has many benefits for parking management in addition to
show promising advances in the field of smart parking user-centric ones. Facility managers can improve parking
technology and promote a future in which parking layouts and overall operational efficiency by using the
inconvenience is minimized through the integration of system's insightful data on parking space utilization
cutting-edge AI technologies. patterns.
II. RELATED WORKS
Keywords: YOLO NAS (you only look once), Python,
vehicle tracking, DeepSORT (deep, simple online real- The authors of the paper "Intelligent Multiple Vehicle
time tracking), Pytorch, vehicle detection. Detection and Tracking Using Deep Learning and Machine
Learning: An Overview," M. Ben Youssef, A. Salhi, and F.
I. INTRODUCTION Ben Salem[1], give a thorough explanation of the function
The increasing density of metropolitan areas and the of autonomous vehicles (AVs) and the significance of
ensuing increase in automobile traffic have created a vehicle classification in smart cities. The authors provide
number of difficulties, one of which is the effective an Autonomous Learning Intelligent cars platform,
management of parking spots, especially in busy highlighting the advancements in autonomous cars (AVs)
commercial settings like malls. One of the common and emphasizing the use of electric vehicles (EVs) as a
problems drivers have in these places is the aggravation means of lowering fuel consumption and noise pollution.
and time they lose trying to find their parked cars in large This platform highlights the application of Artificial
parking lots. This research presents an innovative approach Intelligence (AI) to improve mobility and communication
to this problem by introducing a Smart AI-Based Vehicle among EVs by integrating several physical automobiles
Monitoring System that makes use of the state-of-the-art with an AI-based scheduler for platooning. The study
YOLO NAS (You Only Look Once NAS) object detection explores the tracking and identification of electric vehicles
algorithm. (EVs) at parking lots or stations, suggesting effective
communication tactics and entrance control strategies for
Creative and astute solutions are essential as cities controlling EV traffic. This work provides insights into
struggle to meet the ever-increasing demands on their intelligent vehicle systems and their implications for traffic
parking infrastructure. Our suggested solution offers management in smart cities, which is in line with my
drivers a smooth and convenient driving experience while research focus in AI-based tracking systems for vehicle
resolving the frequent issue of misplaced or forgotten cars monitoring employing YOLO.
in parking lots. Our solution tries to transform parking
The authors of "Ship Detection and Classification in and consistently, guaranteeing that each player's distinct
Cas500-1 Images Based on YOLO Model," T.-H. Kim et identity is preserved throughout the game. This
al. [2], discuss the issues given by the need for effective methodology deviates from conventional filtering.
port operations and ship surveillance, as well as the rising algorithms, demonstrating the application of state-of-the-art
amount of container logistics at ports across the world. deep learning technology for dynamic object tracking. This
Using Compact Advanced Satellite 500-1 (CAS500-1) paper highlights the importance of sophisticated tracking
satellite images, the authors use the YOLOv5 (You Only approaches in dynamic contexts, which is relevant to my
Look Once version 5) deep learning system to recognise research on an AI-based tracking system for vehicle
and categorise a variety of boats, including bulk ships, monitoring employing YOLO
container ships, oil tankers, and small vessels. Using
satellite imagery, the programme attempts to improve ship
Kejriwal et al.'s [6] work, "Vehicle Detection and Counting
monitoring in locations where the Terrestrial AIS (T-AIS)
using Deep Learning-based YOLO and Deep SORT
system is insufficient. Understanding YOLO-based AI-
Algorithm for Urban Traffic Management System,"
based tracking technologies that are relevant to my research
discusses the need of vehicle counting for traffic density
interests in vehicle monitoring.
estimates in the context of intelligent transportation
systems (ITS). The authors propose a video-based vehicle
The authors of the work "Artificial Intelligence (AI) based counting and tracking system to address the inadequacies
Detection of Traffic Violations by Two-Wheeler Vehicles," of current traffic management systems, which routinely
M. Bala Kishore et al.[3], explore the use of AI in detecting disregard traffic density. The system recognizes, counts,
traffic violations that are committed by two-wheeler classifies, and tracks automobiles in real-time by utilizing
vehicles in particular. The authors use artificial intelligence the Deep SORT algorithm and the You Only Look Once
to create a detection system that can track and identify (YOLO) deep learning framework. With datasets tailored
infractions of traffic laws by motorbikes and scooters. The to Indian road settings, the model is trained for a variety of
focus on AI-based detection in the paper is in line with my classes and achieves an average counting accuracy of
research interest in an AI-based tracking system for 86.56% and an average precision of 93.85%. This study
YOLO-based vehicle monitoring, even though the offers important information regarding. real-time traffic
publication's details are brief. This research adds to our management, which is in line with the objectives of clever
understanding of artificial intelligence applications in and effective traffic surveillance systems, and is directly
traffic management and surveillance by shedding light on related to my research on an AI-based tracking system for
possible collaborations between real-time tracking systems vehicle monitoring utilizing YOLO.
like YOLO and traffic regulation enforcement for
In Ahmed et al.'s study "The YOLOv8 Edge: Harnessing
improved monitoring and enforcement.
Custom Datasets for Superior Real-Time Detection,"[7]
Rodrigo Barbosa et al.[4] discuss the significance of they explore the importance of YOLO as a core real-time
intelligent traffic monitoring systems in light of the object detection system with applications in autonomous
emergence of cutting-edge technologies such as artificial vehicles, robotics, and video surveillance. The study finds a
intelligence (AI) and the Internet of Things. The authors vacuum in the body of knowledge on the rapid and accurate
propose a novel approach for vehicle detection called real-time detection of bespoke items. In order to overcome
LightSpaN, which is a sparse and mild version on the this, the authors train a YOLOv8 detector using a specially
CNN-based method. Unlike current techniques, which created dataset, demonstrating how it can transform a
typically display complexity and delay, LightSpaN offers a number of industries. With an overall mAP50 of 0.864 and
quick training model without complicated solutions. The a mAP50-95 of 0.758, the experimental findings show that
Simulation of Urban MObility (SUMO) tool and real-world the specially trained YOLOv8 detector delivers excellent
vehicle traffic implementation with IoT devices are utilised accuracy and real-time performance.
to evaluate the effectiveness of the proposed strategy. The
An et al.'s study, "A Nighttime Vehicle Detection Method
results reveal that LightSpaN outperforms comparable
Based on YOLOv8," addresses the importance of precisely
works in recognising a wide range of vehicle types rapidly
detecting targets in difficult low-light situations, a critical
and correctly. More precisely, especially for emergency
problem in the advancement of autonomous driving
vehicles, an average accuracy of about 99.9% is achieved.
technology. In order to improve target detection
This study emphasises the need of effective and fast
capabilities in difficult settings, especially at night and
training models in real-time traffic monitoring systems,
dusk, the authors highlight the shortcomings of pure visual
which is pertinent to my work on an AI-based tracking
object detection technology in low-visibility environments
system for vehicle monitoring that uses YOLO.
and introduce YOLOv8, a cutting-edge image processing
technology. The study assesses how well YOLOv8
Zhang et al. [5] address the topic of object tracking in recognizes vehicles in low-light conditions and emphasizes
computer vision in their paper "A Sport Athlete Object how much better it performs than earlier YOLO series
Tracking Based on Deep Sort and YOLO V4 in Case of models. The paper highlights the significance of YOLOv8
Camera Movement," focusing on the distinct identity as a state-of-the-art instrument for enhancing autonomous
component, similar to the Re-identification (REID) driving safety in low-light conditions, giving crucial
problem. The study's goal is to create a real-time tracking background for my study on an AI-based tracking system
system for every player in the game by concentrating on for vehicle monitoring using YOLO.
sports-related domains, particularly NBA and World Cup
According to Anvitha et al.'s work, "Analysis of Traffic
scenes. The authors utilize cutting-edge deep learning
Management System using YOLOv8," traffic violation
technology to identify players and the ball by using
detection is a crucial component of modern traffic
YOLOv4 for object detection. After then, the Deep SORT
management initiatives. The study emphasizes the critical
algorithm is used to track the identified objects precisely
role that efficient traffic infraction detection plays in
creating sustainable and flexible cities, highlighting its vision and intelligent transportation systems by offering a
significant influence on traffic management, road safety, viable method for precise and dependable vehicle number
and community well-being. By properly estimating the plate detection.
number of vehicles on a road concurrently and offering
The difficulties in managing parking in cities are addressed
precise information through labeled counts and categories,
in Huang et al.'s[13] paper, "Design and Management of an
the suggested model aids in traffic enforcement and
Intelligent Parking Lot System by Multiple Camera
surveillance. Notably, the model simplifies the process of
Platforms". The writers stress the cost and time-consuming
obtaining thorough vehicle information by identifying
aspects of conventional parking solutions as drawbacks.
instances of cars running red traffic signals. By
Security cameras, micro aerial vehicles (MAVs), and
concentrating on identifying numerous vehicles that break
cutting-edge features like alerting drivers to available
the rules at the same time, the model is able to reach an
spaces, logging parked positions, remote monitoring, and
astounding 97% classification accuracy. This study offers
parking spot leadership services are all included in their
crucial background information for my research on an AI-
planned hybrid intelligent parking system. This work,
powered tracking system that uses YOLO to monitor
which focuses on an AI-based tracking system for vehicle
vehicles in line with the main objective of improving safety
monitoring utilizing YOLO, gives important background
and traffic management.
information for my project. Enhancing vehicle tracking and
monitoring efficiency—especially in urban settings—by
Afonso et al.'s work, "Vehicle and Plate Detection for utilizing Huang et al.'s insights is the goal, in line with the
Intelligent Transport Systems: Performance Evaluation of changing field of intelligent transportation systems.
Models YOLOv5 and YOLOv8,," dives into the topic of
The authors investigate the integration of the YOLO v4
intelligent transportation systems and employs computer
algorithm for intelligent license plate identification in a
vision technology to increase urban mobility efficiency and
smart parking environment in their work, "Automatic
safety. The use of regression-based algorithms, such as
Vehicle License Plate Recognition Based on YOLO v4 for
YOLO (You Only Look Once), for automatically detecting
Smart Parking Management System," by Hsieh et al. [14].
automobiles and licence plates is being researched. The
They stress how important image recognition is becoming,
study evaluates the performance of the YOLOv5 and
especially for applications like autonomous vehicles and
YOLOv8 models using a specifically chosen dataset
factory automation. The importance of license plate
obtained by transfer learning methods. According to the
identification for parking system fee collection and road
study's results, YOLOv8 performed somewhat better than
speed detection is emphasized in the article. Using
YOLOv5, with an accuracy rating of 97.19% and a
automatic license plate recognition and LINE functions, the
precision rating of roughly 97.98%. Training time is
YOLO v4 algorithm—which is renowned for its
another factor that works in YOLOv8's favor. These
convolutional neural network architecture—is used to
revelations enhance my contextual comprehension of study
create an intelligent parking management system. In the
on a YOLO-based AI-based tracking system that highlights
campus environment, the experiment shows a noteworthy
the effectiveness of the models in intelligent transportation
100% identification rate for license plates. This study offers
applications.
insightful background for my research on an AI-based
tracking system that uses YOLO to monitor vehicles,
demonstrating the algorithm's adaptability for use in smart
The research "Robust Licence Plate Recognition With
transportation applications.
Shared Adversarial Training Network," by Zhang et al.
[11], examines the difficulty of recognising licence plates III. METHODOLOGY
in a variety of challenging contexts. The study proposes a
Shared Adversarial Training Network (SATN) in
acknowledgment of deep learning's contribution to licence
plate recognition (LPR). This network efficiently identifies
environment-independent along with perspective-free
semantic characteristics from wild licence plates by
utilising its prior knowledge of normal stencil-rendered
licence plates. Moreover, a novel Dual Attention
Transformation (DAT) module is proposed to correct
highly perspective-distorted license plate features. In the
LPR task, the proposed method outperforms the state-of- Fig1: Block Diagram of License Plate Detection Using
the-art techniques, according to extensive testing on DeepSORT and YOLO NAS
AOLP-RP and CCPD benchmarks. This study provides
insightful information about improving the robustness of A. Dataset Description
license plate recognition technologies, which set the stage The 433 photos in the Car License Plate Detection
for my research on an AI-powered tracking system that dataset, which was obtained from Kaggle, have bounding
uses YOLO to monitor vehicles. box annotations indicating the location of the car license
A method for vehicle number plate detection using plate. There are annotations available in PASCAL VOC
YOLOv8 and EasyOCR is proposed by Patil et al. [12]. format.
The study uses EasyOCR for effective optical character B. YOLOV8
recognition and leverages the YOLOv8 model for accurate
The initial step involves leveraging libraries such as
vehicle number plate detection. These technologies work
OpenCV or Pillow to proficiently load images and
together to improve automated vehicle number plate
annotations from the dataset. This process encompasses
detection systems' accuracy and efficacy. The paper, which
reading image files, commonly stored in formats like JPEG
was published in 2023, advances the field of computer
or PNG, and parsing annotation files, often structured in video streams. It analyzes visual data using models and
formats such as PASCAL VOC XML. Following data algorithms to identify the presence, location, and frequently
loading, a comprehensive exploration of the dataset the kind of vehicles. YOLO (You Only Look Once) and
becomes imperative. This exploration phase encompasses Faster R-CNN (Region-based Convolutional Neural
key tasks like determining the total number of images, Network) are two popular object identification frameworks.
inspecting their dimensions, and scrutinizing the format of Applications for vehicle detection include autonomous
the annotations. Additionally, researchers will delve into vehicles, surveillance, and traffic monitoring. By giving
visualizing sample images alongside their associated real-time information regarding the presence and
bounding boxes, providing an insightful overview of the movement of vehicles inside a specific area, it plays a
data distribution and potential challenges that may arise significant role in improving road safety, traffic
during model training. This meticulous data loading and management, and overall transportation efficiency.
exploration phase sets the stage for a thorough
understanding of the dataset's characteristics, ultimately IV. RESULT ANALYSIS
informing subsequent steps in the development of the AI-
based parking management system. A comprehensive assessment strategy is presented. The
C. Vehicle Tracking system's performance is rigorously evaluated utilizing
metrics such as detection accuracy, tracking accuracy, and
Vehicle tracking involves real-time monitoring of identification accuracy. This quantitative analysis provides
vehicle locations using GPS and telematics devices. insights into the system's overall efficacy in detecting and
Telematics devices collect data on speed, acceleration, and tracking vehicles within parking scenarios. Additionally,
fuel consumption. These systems provide real-time location the parking data is subjected to a detailed analysis aimed at
updates through web platforms or mobile apps, allowing identifying peak parking times and optimizing space
users to visualize routes on digital maps. Historical data allocation during high-demand periods.
analysis aids in optimizing routes and improving
efficiency. Alerts can be set for events like speeding or
unauthorized use, and integration with other systems
enhances functionality. Vehicle tracking finds applications
in fleet management, logistics, public transportation,
security, and personal vehicle monitoring, offering benefits
such as improved efficiency, security, and logistics Fig2: Qualitative Results of Number Plate detection using
optimization. YOLO NAs
D. DeepSORT The system's capability to detect unauthorized parking
DeepSORT (Deep Simple Online and Realtime violations is scrutinized, contributing to enhanced security
Tracking) is an advanced object tracking algorithm that measures. Furthermore, the generation of detailed reports
extends the capabilities of the SORT algorithm by for parking management and security personnel is
integrating deep learning techniques. It excels in real-time emphasized, offering valuable insights derived from the
tracking scenarios within video streams, offering improved AI-based monitoring system. This multifaceted evaluation
accuracy and robustness. DeepSORT incorporates a deep and analysis approach not only gauges the immediate
appearance descriptor model, pretrained on extensive performance metrics but also establishes the practical
datasets, to enhance object differentiation, especially in utility and effectiveness of the proposed system in
crowded or occluded scenes. This deep learning aspect optimizing parking management and ensuring security
contributes to superior identity preservation across frames, within the monitored environment.
making DeepSORT effective in complex tracking
scenarios. Widely applied in surveillance, autonomous
vehicles, and sports analytics, One notable feature of
DeepSORT is its capacity to deliver precise and
dependable object tracking in dynamic situations.
E. Pytorch
One popular open-source machine learning library for
deep learning applications is PyTorch. Created by
Facebook, it provides a dynamic computational graph,
enabling intuitive and flexible model development.
PyTorch supports dynamic computation, making it favored
for research and experimentation. Its extensive community
and documentation contribute to its popularity. With a user- Fig3: YOLO V8 Model Performance on
friendly interface, PyTorch simplifies the creation and Precision, Recall, and Loss at Different Training Epochs
training of neural networks. It offers seamless GPU
acceleration, facilitating efficient model training. PyTorch's During model training, considerable gains are noticed
dynamic nature, eager execution, and TorchScript for across a variety of criteria. The train/box_loss,
model deployment make it a powerful framework for both train/cls_loss, and train/dfl_loss values constantly fall,
research and production-level machine learning tasks. showing improved performance when predicting bounding
boxes, classifying objects, additionally predicting object
F. Vehicle Detection direction, focus, and position. On the validation set, the
A computer vision approach called "vehicle detection" model shows generalisation by reducing val/box_loss,
is used to locate and identify automobiles in pictures or val/cls_loss, and val/dfl_loss. Furthermore, measures like
as precision(B), recall(B), mAP50(B), and mAP50-95(B) Box object Cls
Epoch Precision Recall Map
progressively improve, demonstrating the model's capacity loss loss loss
to generate correct predictions, properly find objects, and 10 0.1 0.6 0.1 0.0055 0 0.81
attain improved precision and recall rates. The overall trend 20 0.1 0.65 0.1 0.005 0 0.8
indicates that the model improves its object identification 30 0.2 0.55 0.09 0.004 0 0.79
capabilities and generalises effectively to previously 40 0.3 0.44 0.08 0.004 0 0.78
unknown data throughout the training period. 50 0.4 0.55 0.08 0.0035 0 0.77
60 0.5 0.45 0.06 0.002 0 0.76
70 0.5 0.45 0.05 0.003 0 0.75
80 0.6 0.34 0.04 0.0025 0 0.76
90 0.7 0.34 0.03 0.0013 0 0.77
100 0.8 0.45 0.02 0.0015 0 0.78

Effective object detection and few false positives are

indicated by metrics like precision and recall, which
continuously hold strong at 0.96 during validation. The
model's mean average precision (mAP) also shows
improvement, with mAP at 50% IoU increasing from 0.81
Fig4: YOLO NAS Model Performance on Precision, to 0.9775, and mAP at 50% to 95% IoU rising from 0.5 to
Recall, and Loss at Different Training Epochs 0.8, underscoring the model's enhanced precision in
The model shows steady improvement across the training detecting both regular and challenging objects, even when
phase in all kinds of criteria. The train/box_loss, small or occluded.
train/cls_loss, and train/dfl_loss, in particular, all show a V. CONCLUSION
consistent decline, suggesting improved ability to forecast
bounding boxes, classify objects, and predict the direction,
focus, and placement of objects. The model's generality is The effective use of YOLOv8 in the AI-based car tracking
demonstrated on the validation set by declining system is evidence of a methodical and thorough
val/box_loss, val/cls_loss, and val/dfl_loss. Furthermore, methodology. Selecting YOLOv8, which is well-known for
metrics like accuracy(B), recall(B), mAP50(B), and its real-time object identification capabilities, along with
mAP50-95(B) gradually rise, indicating that the model is advanced tracking algorithms has resulted in an excellent
becoming more accurate at predicting outcomes, locating system performance, exhibiting remarkable parameters
objects efficiently, and achieving higher recall and including F1 score, recall, accuracy, and precision.
precision rates. All of these tendencies point to the model meticulous preprocessing techniques include data
improving over the training period in terms of item normalization and augmentation are essential to improving
detection and successfully generalizing to new data. the model's flexibility in a variety of driving scenarios and
environmental circumstances. These preliminary actions
strengthen the system's resilience and enable it to manage
real-world situations with effectiveness. The strategic use of
post-processing, particularly Non-Maximum Suppression,
stands out as a key element in refining the model's
predictions. By minimizing false positives and improving
bounding box accuracy, the system's reliability is
significantly enhanced, ensuring more accurate and precise
object tracking. The deployment phase highlights the
system's versatility by successfully optimizing the YOLO
NAS model for real-time processing on different platforms.
Fig5: YOLO V5 Model Performance on The emphasis on ongoing monitoring and maintenance
Precision, Recall, and Loss at Different Training Epochs underscores a proactive approach to guarantee the system's
The model performs brilliantly during training on a variety adaptability to changing conditions and to uphold accuracy
of criteria. Train/loss, a indicator of overall training loss, over time. User interaction is recognized as a crucial aspect,
rapidly decreases from 0.15 to around 0.015, demonstrating facilitated through a user-friendly interface that provides
the model's consistent learning and improvement in precise real-time tracking information. This feature enhances the
prediction-making. More precisely, train/box_loss and practicality and usability of the system, making it more
val/box_loss fall from 0.015 to 0.01 and 0.0045, accessible and user-friendly in real-world applications.The
correspondingly, in bounding box prediction, suggesting acknowledgment of continuous testing and iteration
higher accuracy and generalisation to new data. As losses emphasizes the dynamic nature of the system. This
decrease from 0.04 to 0.00 during training and remain commitment to improvement and adaptability positions the
constant at 0.02 on the validation set, object classification vehicle tracking system as a reliable solution capable of
competence, as represented by train/cls_loss and addressing evolving challenges across diverse monitoring
val/cls_loss, considerably increases, proving the model's
scenarios. In summary, the implemented AI-based vehicle
capacity to reliably identify items and its generalizability.
tracking system, with its careful selection of YOLOv8,
TABLE1. YOLO NAS Model Performance Metrics Over meticulous preprocessing, effective post-processing, and
Training Epochs user-centric design, stands as a testament to its high
accuracy, adaptability, and utility in various real-world
applications. The system's success is not just a result of a
one-time implementation but reflects an ongoing
commitment to monitoring, maintenance, and continuous
improvement.