Smart Traffic Surveillance: An AI-Based System for

Real-Time Accident Detection and Emergency

Notification
Raymond B. Sedilla Jaceb Vinzreinier S. Gile Franchesca Marielle N. Giron Marie Louise Z. Lariosa
School of Information and School of Mechanical School of Civil, Environmental, School of Mechanical
Technology, and Manufacturing Engineering, and Geological Engineering, and Manufacturing Engineering,
Mapua University, Mapua University, Mapua University, Mapua University,
Manila, Philippines Manila, Philippines Manila, Philippines Manila, Philippines
rbsedilla@[Link] jvsgile@[Link] fmngiron@[Link] mlzlariosa@[Link]

Abstract — With the rising number of road accidents in urban rates per 100,000 population, being safer than 8 other
areas, traditional traffic surveillance systems have proven Southeast Asian countries—with the first having 150% higher
insufficient due to their reliance on manual monitoring and fatality rates in comparison [1]. In fact, even with the addition
delayed emergency responses. This study introduces a conceptual
of COVID-19, road injuries ranked 14th from 139 cause of
framework for an AI-based traffic surveillance system designed
to detect vehicular accidents in real-time and automatically notify
mortalities in the Philippines. In addition, the report has also
emergency services. By integrating deep learning, object stated that almost 36 fatalities occur every thousand of
detection, and motion analysis into existing CCTV infrastructure, kilometers on the road. This implies that, in recent years,
the proposed system aims to distinguish between normal traffic there’s an increasing trend with road casualties, and almost 25
flow and collisions with greater speed and accuracy. It also road-related deaths occur every day. In addition to that, each
considers performance under challenging conditions such as low of those accidents imposes a significant economic burden. In
light or adverse weather. Through a comparative, literature- fact, the total economic cost from the total road accidents
based analysis, the study demonstrates the potential of AI to representation is estimated to be 11 billion USD, which is
transform road safety management and contribute to the
equivalent to about 3% of the country's total GDP. These
development of more responsive, data-driven, and intelligent
urban transport systems. emphasize that road safety remains a national welfare issue,
and highlights the need for a well-developed road safety
Keywords — Artificial Intelligence (AI), Real-Time management system.
Accident Detection, Smart Traffic Surveillance, Emergency
Notification System, Deep Learning, Intelligent Transportation
Systems (ITS)

I. INTRODUCTION
1.1. Research Background
1.1.1. Road Accidents and Hazards
Several news articles regarding road accidents are
being published every day, and it is becoming a critical
concern not only for the economy, but also the welfare of
people living inside the community. People, such as students
and commuters, are afraid of being in a random road incident,
whether as the witness or the victim. One reason is that they
are usually not equipped with knowledge on how to respond
properly in such situations, and this somewhat lowers the
victim’s chance of survival. Secondly, they are also afraid as
they are not perfectly secured on rescue operation response
systems as those incidents occur. That is because human
operators of the current surveillance system are bound to
overlook road accidents, since they are not expected to
monitor live video feed all day. With a rapidly increasing
number of vehicles on the road per unit area, death statistics Figure 1.1.1. Philippines Road Safety Profile 2021 (ATO,
are getting correspondingly higher. As reported in the 2025)
Philippines Road Safety Profile, the country has 10 fatality

1
 Accordingly, the notable factors that contribute to surveillance utilizes features such as road accident detection
fatalities in roads consist of low-quality road infrastructure, and post-crash emergency response. Instead of the traditional
lack of education about road safety, driver’s behavior, and lax surveillance system which commonly has delays with
enforcement of traffic laws [2]. This serves as the main emergency response time and ineffective risk management, the
framework that enables a good strategy for improving the road automated surveillance system enables a capable and better
safety system. The common strategies being discussed include real-time response and analysis, which could further improve
improving the quality of the road infrastructure system, safety in roads.
advocate road education and awareness, develop better driver
training programs, enforce stricter legislative framework, and 1.2. Significance of the Study
quicker emergency response. While these strategies are Road accidents continue to be a serious issue,
commonly promoted, they may not be the most effective on especially in busy urban areas where traffic congestion and
their own because they rely heavily on long-term policy limited emergency response times can make the situation
changes, human behavior, and infrastructure investment, worse. While many roads are already equipped with
which often take time and great resources to implement and surveillance cameras, these systems often fall short when it
sustain. Hence, some research tackles different solutions, and comes to spotting accidents quickly or alerting the right
one of its prospective approaches is to integrate artificial authorities in time. This study seeks to address that gap by
intelligence to the existing traffic surveillance system. introducing a smart traffic surveillance system powered by
Correspondingly, AI-enabled surveillance systems help detect artificial intelligence (AI) that detects accidents in real-time
accidents and even initiate rescue and emergency operations. and sends immediate alerts to emergency services. Such a
In any case, AI-integrated traffic surveillance systems have system holds the potential to significantly reduce response
potential impact on safety road management, ultimately time, thereby minimizing loss of life and the severity of
alleviating the risks posed by road accidents and hazards. injuries.

1.1.2. Traffic Surveillance Cameras From a technological perspective, the study

Surveillance systems have been playing a crucial role highlights how AI can enhance existing infrastructure,
in improving safety, especially in the roads. Over time, Closed particularly CCTV networks, by enabling real-time traffic
Circuit Television (CCTV) cameras are undergoing rapid analysis and autonomous decision-making. Traditional
notable transformations that continuously impact securities in systems are limited to passive monitoring and often
the system. From its role as public safety for the police force, underperform in adverse conditions like poor lighting or
to its use as a traffic surveillance camera for the transport weather. By applying deep learning algorithms and computer
system. Various features are being added to the closed-circuit vision models, the proposed system can recognize abnormal
television (CCTV) such as license plate and speed analysis. vehicle behavior and trigger emergency notifications instantly.
With the advancement of artificial intelligence, traffic It also aligns with broader smart city goals, where intelligent
surveillance is also expected to evolve alongside the machine transport systems (ITS) aim to create safer, more efficient
learning algorithms, which could lead to the new smart traffic urban environments through data-driven technology.
surveillance systems.
Academically, this research contributes to the
1.1.3. Smart Traffic Surveillance System growing field of AI-enabled road safety systems by
As the concepts of Artificial Intelligence are being synthesizing insights from transportation engineering,
explored, many of its potentials are continuously being applied computer science, and emergency response management. It
to various areas such as crowd control, emergency response, provides a theoretical framework for using object detection,
and safety management systems. Artificial Intelligence motion analysis, and real-time video analytics in accident
enables a transformative approach to road safety, with its detection and offers a literature-based comparison with
potential application in Advanced Driver Assistance System conventional systems. These findings serve as a reference for
(ADAS), smart traffic lights, real-time alert and risk mitigation future studies in intelligent surveillance, helping guide
response, and driving behavior monitoring system. That is the innovation in both public policy and technical design.
most basic use of AI, and it is still constantly improving over Ultimately, the study reinforces the role of AI as a
time. In fact, The UN Agency reports that artificial transformative force in building safer and smarter
intelligence enables the global initiatives to aim to utilize AI to transportation systems.
save halves of death due to road crashes by 2030 [3]. This
suggests that even administrations are quite hopeful about the 1.3. Objectives of the Study
future of AI-integrated systems for road management. This study aims to develop a conceptual framework
for a smart traffic surveillance system that utilizes artificial
Specifically, a smart traffic surveillance system is an AI- intelligence (AI) to detect road accidents in real-time and
integrated surveillance system which uses all kinds of features automatically notify emergency services. The primary goal is
of AI to improve overall road safety. This smart traffic to enhance the current limitations of traditional surveillance

2
systems by introducing automated and intelligent video traditional road surveillance and emergency response in
analysis capable of identifying vehicular collisions as they contrast to the proposed framework. Testing and validation
happen. This involves leveraging object detection and motion will be done through both qualitative and quantitative means
analysis techniques to distinguish between normal traffic flow, involving the satisfaction of respondents on current road
congestion, and critical incidents such as accidents. monitoring systems and theoretical literature-based systems
comparing the effectiveness and feasibility of existing
Another key objective is to explore and integrate AI methods. The limitation will be based on its conceptual and
algorithms, including deep learning models, that can research-based nature since no actual coding will be integrated
accurately process video feeds from existing CCTV in the framework.
infrastructure. The system is expected to analyze various
traffic scenarios and make timely and reliable predictions
about potential road hazards. In doing so, the study aims to II. LITERATURE REVIEW
design a notification mechanism that automatically generates In this chapter, related literature and studies were reviewed
emergency alerts and transmits them to appropriate response to identify the limitations of traditional traffic surveillance
units, reducing the delay between accident occurrence and systems and to explore how artificial intelligence and deep
medical intervention. learning offer improved solutions for road safety. Specifically,
the aim is to establish a foundation for the present study by
In addition, another purpose of the study is to conduct examining past approaches and recent advancements that
a comparative analysis using related literature and existing discusses the development of the smart traffic surveillance
studies to identify the most effective AI-based traffic systems.
surveillance models. This involves evaluating each system
based on its methodology, focusing on the processes and 2.1. Traditional Traffic Surveillance System Challenges
techniques. By analyzing how different studies approach the One of the key aspects in understanding the
design and implementation of their systems, the research aims limitations of traditional traffic surveillance systems lies in
to determine the gaps between similar works and highlight determining how its fundamental components operate. These
areas for improvement. This comparison will guide the components involve traffic monitoring system approach, video
development of an optimal process and system design that surveillance system, and hardware limitations.
effectively addresses the limitations of traditional surveillance
methods. 2.1.1. Traffic Monitoring System Approach
The traditional video surveillance system
Additionally, the study seeks to evaluate the system’s approach focuses on creating more efficient real-time
performance in different environmental conditions such as traffic management to alleviate concerns of traffic
nighttime, low visibility, or heavy rain, which are factors that congestion and road safety. Despite the potentials of
often compromise the effectiveness of traditional surveillance. the surveillance system to improve security, the lack
A comparative analysis will also be conducted to highlight the of modern automated features lessens its impact to
differences between manual and AI-assisted monitoring in road safety. This is due to the video cameras being
terms of response time and detection accuracy. Lastly, this limited to passive monitoring tasks and weak
research intends to assess the feasibility and limitations of automated processing [4]. As a result, the traditional
implementing such a system conceptually, using a literature- systems struggle to detect and respond to road
based framework without actual coding that focuses on its accidents in real time.
potential impact on road safety and smart city development.
Consequently, the lack of advanced video
1.4. Scope of the Study processing systems for surveillance cameras
This study focuses on developing a smart traffic highlights the need for manual operators. Traditional
surveillance system framework that utilizes artificial surveillance systems are being managed by human
intelligence (AI) to detect road accidents in real-time and operators by sitting in control rooms filled with the
automatically notify emergency services. The framework is cameras. However, this method of surveillance is
designed using existing literature and evaluated through a slow, error-prone, and inefficient [5]. Accordingly,
conducted comparative analysis of various studies. The study such limitation implies the need for an alternative
focuses on the effectiveness of integrating AI algorithms in monitoring approach than manual observations as a
existing roadside cameras, specifically object detection and surveillance system to reduce human-errors and
motion analysis, in identifying abnormal vehicle behavior and delays, ultimately leading to improved safety in
roadside collisions. The geographic scope of the study is transportation systems.
limited to urban intersections and highways where
surveillance cameras are readily available to produce 2.1.2. Video Surveillance System
immediate results. The study also explores the comparison of

3
 Aside from the limitations in the past it advances. Currently, AI-integrated surveillance
approach which required human operators, the systems have found its application in transportation
traditional video surveillance system is hindered by safety management. Artificial intelligence can
lack of its functional capability. The lack of system’s manage and optimize traffic lights, form personalized
scalability reduces the total possible numbers of route systems, and enable vehicles with AI-supported
camera inputs. In addition, it also has limited features for a safe transport system. However, some
retention capability, which also implies storage drawbacks with regards to ethical and legal
capacity for relevant footage that could potentially be challenges such as the AI system requires substantial
valuable data [6]. These functional limitations hinder amounts of information about road users, and
the flexibility of the system and highlight the need for transparency and accountability from the legal
more advanced solutions tailored to enhancing road framework of AI usage [10]. In addition, due to the
safety. increasing number of vehicles and drivers running at
high-speeds, road accidents frequency also increases.
2.1.3. Hardware Limitations of Surveillance Systems To alleviate the issues with road crash and pedestrian
Lastly, there are numerous hardware accidents, surveillance cameras are integrated with
limitations on traditional surveillance cameras. These automated accident detection and response. The IoT
hardware properties are termed the field of view, kit is utilized for accident prevention, and deep
video and image resolution, dynamic range, learning-based models are essential to validate its
durability, audio volume, and privacy [7]. The output to ensure it is not a false alarm and initiate
traditional surveillance cameras such as CCTVs, have rescue operation [11]. This combination enhances the
low resolution and dynamic range, which is vital for reliability of traffic surveillance systems, making
video processing and analysis. Additionally, the them more responsive and effective in preventing
traditional camera's properties are susceptible to fatalities and minimizing the impact of road
weather conditions or environmental changes. The accidents.
reliability of alarm systems is being factored by the
adverse weather conditions and low-illumination 2.2.2. Deep Learning
environment [8]. This implicates the need for the Deep learning is a more advanced feature of
surveillance camera’s hardware properties to also be machine learning as it can analyze more complex
compatible for AI features such as detection and data ranging from images to footages. This is more
emergency response systems. suitable for images and videos as it is constructed
with artificial neural networks to model after the
2.2. AI in Traffic Surveillance and Monitoring human brain’s architecture and way of operation.
Traffic monitoring had always been a vital part and Correspondingly, most of the surveillance system
important effort to managing road systems and infrastructures, utilizes video footage and images to generate locally
as well as transportation engineering. Yet, the lack of relevant road statistics. The crash and accident
automated response of surveillance systems warrants the need detection systems were making progress by
for human-centered operators. These could lead to challenges increasing the algorithm’s capability to read details
in monitoring management in terms of inaccuracies due to from images real-time. The datasets utilized for
human-error, inconsistent tracking, and wasteful time developing the deep learning algorithm’s ability to
consumption. However, the rapid advancements of artificial analyze crash accidents came from online sources
intelligence, deep learning, and computing widen the scope of references, where it represents real-world scenarios of
video-based traffic monitoring systems [9]. Artificial accidents, regular traffic behavior, and near misses
intelligence could alleviate all those challenges by enabling [12]. Various researchers develop all kinds of deep
automatic video analytics and processing that could learning architecture and algorithms to improve the
summarize relevant data over the entire day of operation with overall processing, response, and predictive analytics.
consistency and accuracy. Such details involve traffic Predictive analytics enables deep artificial
congestion, license plate identification, speed tracking, and intelligence to identify patterns and predict future
accident prevention which are essential to forming good outcomes. One of the research projects utilizes Real-
strategies and approaches to transportation management. Time Collision Avoidance (RTCA) through deep
learning architecture to develop a prediction
2.2.1. Road Safety Management capability [13]. This RTCA enables real-time object
Artificial intelligence was utilized primarily identification, instantaneous position tracking, and
on road management to improve overall safety, aside distance estimation. These attributes lead to its
from its purpose of organizing relevant data from the capability to predict the possibility of the accidents
roads. This continuously transforms traffic within 0.09 seconds. However, the main problem
management as it offers new opportunities each time with AI in crash detection lies with false positive

4
 alarms and accuracy. As deep learning and artificial are also existing IoT-based networks for railway accident
intelligence continues to advance, they hold greater prevention systems that can also be integrated with road
potential to enhance the accuracy and reliability of accident detection, such as the ESP8266 which monitors track
traffic surveillance which could contribute to safer conditions and human activity; this reduces accidents in
roads through more effective crash detection and railway crossings [20].
prevention systems.
2.4. Machine Learning for Collision Recognition
2.2.3. Real-Time Incident Detection A vehicle collision detection method utilizing
Aside from car crashing accidents, machine learning integrates supervised and unsupervised
pedestrian detection and tracking are also an algorithms to enhance accuracy and recall rates. It processes
important factor that affects total road incidents. various vector data to calculate collision probabilities and
Many researches about accident detections systems classify input vectors as collision or non-collision [21]. This
are proposed but they mostly have a trade-off issue shows that machine learning is adaptive for accident
between accuracy and speed of detection. Hence, a prediction whilst using position mechanisms for unusual
reliable real-time surveillance system achieves an vehicle movement detection. Another machine learning
optimal balance based on its accuracy, speed, and system operates with the K-means expansion algorithm for
high requirements of computational power [14]. intelligent data collision. This algorithm enhances precision
Some research involves the use of YOLOv11 to and speed in recognizing table and field associations, even
propose a highly efficient real-time detection and with non-uniform data sources and low-end hardware,
emergency response [15]. This model achieved a reducing manual labor and costs [22].
Recall of 0.8249, an F1-Score of 0.9040, and
maintained a perfect Precision score of 1.0000. This Machine learning techniques also utilize the Internet
implies that although only 82.49% of positive of Vehicles (IoV) for collision detection, achieving a 94.4%
accidents are recorded, all of those were accurately accuracy in predicting forward collisions. It emphasizes the
identified. However, real-time accident detection is significance of various parameters influencing traffic
still continuously being improved in order to account accidents, contributing to enhanced road safety [23]. A vehicle
for perfect balance of response time, speed, accuracy, recognition method using machine learning analyzes multi-
and determining all positive crashes amidst ever line point cloud data from laser radar, identifying vehicles and
changing weather conditions while minimizing sending alarm information to prevent property and life damage
hardware limitations. during potential collisions [24]. There is also an existing
systematic approach for machine learning algorithms such as
2.3. IoT Based System for Accident Prevention the Multilayer Perceptron (MLP). This technique is effective
Recently developed technologies utilize the for detecting potential collisions in Highly Automated Driving
integration of Internet of Things (IoT) based algorithms and systems, aiding in hazard analysis and risk assessment during
software in order to improve safety transportation systems. the safety-related concept phase [26].
Traditional traffic systems mostly rely on human input and
static equipment which does not provide accurate results [16]. 2.5. Environmental Adaptability for Detection System
Compared to traditional systems, IoT systems work faster and Maintaining surveillance accuracy under varying
can give more helpful information. The anatomy of this environmental conditions remains a critical challenge for both
system consists of simultaneously working sensors and traditional and AI-enhanced traffic monitoring systems. Real-
machine learning algorithms to produce better results in world environments expose surveillance cameras to
accident detection. IoT-based Smart Communication Systems unpredictable factors such as heavy rain, fog, glare, and
(SCS) are primarily built for active road monitoring. The nighttime darkness, which reduce the clarity of video feeds
communication feature of this network functions to alert and limit the effectiveness of object recognition. These
nearby vehicles and emergency units that an accident has conditions lead to delayed or failed detection of road incidents
occurred. Utilizing this system is highly advantageous for —especially when traditional CCTV infrastructure lacks
secondary collision prevention in high-traffic areas [17]. advanced imaging capabilities such as infrared sensors,
thermal cameras, or wide dynamic range lenses. Such
Other machine learning applications of IoT-Based limitations compromise real-time detection and classification
systems include Random Forest and LightBGM, which are accuracy, particularly when the visual scene is noisy or
utilized for accident prediction in terms of changing weather unclear [26]. These vulnerabilities point to the need for a
conditions. LightBGM has successfully predicted 2.8 million smarter approach that integrates adaptive visual processing
accidents with a 72% machine learning accuracy [18]. IoT is with reliable object recognition.
also used to develop driver monitoring systems that enable the
application of sensors which detect blinking patterns to AI-based systems offer promising solutions to
analyze driver fatigue and exhaustion [19]. Relatively, there overcome these environmental constraints through deep

5
learning models that are trained on diverse datasets, which do analyze data continuously, only activating response protocols
capture traffic scenarios under multiple visibility levels and when anomalies are detected. One such system uses CCTV
weather conditions. An AI-enabled accident detection and analytics to detect incidents and immediately initiate
alert system using IoT and deep learning proved effective notifications to relevant agencies [30]. Their work shows how
across different environmental contexts, demonstrating how combining video analytics with communication protocols can
environmental variability could be mitigated by training bridge the gap between detection and response, ensuring that
models on weather-diverse datasets to improve detection emergency services are informed with minimal lag. This
consistency [11]. A system called “Heimdall” was also aligns with the goal of reducing the average response time to
introduced—an infrastructure built for anomaly detection that traffic incidents, which in turn can significantly lower fatality
adjusts its interpretation based on contextual traffic behaviors risks.
and illumination changes [28]. These features allow AI
systems to detect incidents in scenarios where traditional Several studies further emphasize the role of IoT
systems might fail, like low-light road segments or during integration in enhancing the responsiveness of these
extreme weather. frameworks. A comprehensive review of IoT-based accident
detection systems found that combining AI with sensors, GPS
In addition to software adaptability, hardware data, and cellular connectivity allows the system to
improvements are also key to addressing environmental automatically transmit coordinates, impact severity, and
challenges. Integrating thermal vision, high-resolution sensors, potential casualties to responders [31]. This level of contextual
and multi-angle camera placements allows for better visual detail is crucial for prioritizing emergencies and allocating
data collection in difficult conditions. This hybrid approach of medical resources [Link] study argued that an
combining AI with upgraded surveillance hardware has been advanced AI-driven emergency response system must include
shown to enhance robustness in real-world deployments. Real- both incident classification and adaptive alert generation to
time incident detection systems in smart cities must be able to ensure that critical events are not overlooked in favor of less
function regardless of environmental disruptions in order to urgent ones [32]. When paired with centralized dashboards or
maintain rapid response and operational continuity [28]. These mobile applications used by traffic operators and rescue teams,
findings are supported by the idea that a well-functioning these systems offer a seamless chain of detection, validation,
emergency response system depends not just on AI accuracy and action.
but also on its adaptability to external factors such as lighting
and weather [29]. As smart city infrastructure continues to Ultimately, an efficient emergency notification
evolve, environmental adaptability will become a fundamental framework must integrate multiple technologies, such as deep
design requirement for effective traffic surveillance, ensuring learning, IoT, cloud communication, and real-time analytics,
round-the-clock monitoring regardless of external conditions. to ensure a reliable flow of information from the site of the
incident to those tasked with responding. A systematic review
2.6. Real-Time Emergency Notification and Response noted that the most successful smart emergency systems are
Framework those that minimize manual handling, reduce latency, and
The ability to not only detect but also immediately ensure data-rich alerts [29]. These findings support the need
respond to road accidents is a critical feature of any smart for future road safety technologies to focus not only on crash
traffic surveillance system. While accurate detection is detection accuracy but also on enabling fast, intelligent, and
essential, the true value of such systems lies in how swiftly data-driven emergency responses.
and effectively they relay critical information to emergency
services. Delays in this process can be life-threatening,
particularly in high-impact collisions or areas with low foot III. RESULTS AND DISCUSSION
traffic where incidents may go unnoticed by bystanders. To In this chapter, the findings will be discussed
address this, researchers have focused on the development of corresponding to the potential impact of the research.
AI-powered frameworks capable of generating real-time alerts Specifically, this aims to elaborate findings in terms of
once a crash is detected. One proposed model integrates product’s optimal design and process, application, and
incident detection with automatic alert systems, enabling evaluation in comparison with existing traffic surveillance
emergency units to receive information within seconds of the systems.
crash event. The system emphasizes the need for automation
in the notification process, reducing reliance on human A. Optimal Process and Design for Smart Traffic
intervention and speeding up emergency deployment [28]. Surveillance System
YOLO designs are typically used for the analysis of
Artificial Intelligence enhances this process through road accidents and car crashes. It enables an algorithm for
its ability to process live video streams, identify abnormal real-time object recognition that is exceptional in its response
events, and trigger notifications without delay. Real-time alert and object detection simplicity. It utilizes the datasets that
frameworks often make use of event-driven architectures that came from the internet, and this resolves the problem with

6
lack of data coming from real-time accident [Link] are
various versions of YOLO, and each has its own different Test Set 1800
features and advantages. Comparative analysis had been made
in order to analyze the effectiveness of each version to detect
accidents accurately [15]. Specifically, the analysis has been Table 3.1.1. Distribution of Data for Training, Testing,
conducted through YOLOv9, YOLOv10, and YOLOv11 with and Validation (Arefin, 2025)
a similar number of parameters of about 25 million, to ensure
fairness in its comparison result. Furthermore, the training involves the use of Google
Colab GPU to ensure efficient computation and results.
Through this application and adapting to 100 epochs level, the
YOLO models were each trained with converted images of
640 pixels for its consistency in reading. Also, pre-trained
YOLO models are still fine-tuned so they can detect accidents
more effectively. To make the models more reliable, training
used data augmentation methods like scaling, cropping, and
flipping images. Important settings like learning rate and batch
size were also adjusted to improve results. The training was
Figure 3.1.1. Accident Detection Process (Arefin, 2025) carefully matched with their dual-head design, which helped
them detect accidents more accurately in complex situations.
For annotation of the dataset images from the
internet, the Roboflow platform is utilized. Roboflow features
extreme analytic capability in relation to vision-based data
such as images and videos. It is also highly efficient and helps YOLO Recall Average F1-Score Average
improve the process of annotation, data collection, processing, Version Precision Latency
and model training. This platform is very important as it (AP) (GPU)
enables the YOLO model to analyze and manage complex
visual data in real-time efficiently, which is vital for object
and accident detection. These vision-based datasets were YOLOv9 0.8209 0.9942 0.9016 20.33ms
divided into three categories: train, validation, and test sets to
properly assess the reliability and performance of the YOLO
model. At a total vision-based data of 9,000 images, the train YOLOv10 0.8102 0.9961 0.8951 20.03ms
set, valid set, and test set have corresponding distributions of
6,310, 890, and 1800, respectively. The train set is utilized to
improve the capability of the system to learn patterns and
features regarding the dataset. Validation set tunes the model YOLOv11 0.8249 0.9940 0.9040 19.93ms
amidst the algorithm development to ensure that its
performance remains accurate even in new data outside the
training set. The test set is utilized after training to provide a Table 3.1.2 Performance Matrix of Different YOLO Model
final, unbiased evaluation of the model’s performance on truly
unseen data, reflecting how well it will work in real-world Among the three YOLO versions evaluated,
situations. YOLOv11 demonstrates the best overall performance. It
achieves the highest F1-Score of 0.9040, indicating a strong
balance between precision and recall, and the highest recall
Total Classes 1 (0.8249), meaning it detects the most accident instances
among all three. While its Average Precision (0.9940) is
slightly lower than YOLOv10’s (0.9961), the difference is
minimal and does not outweigh the advantages in recall and
Total Images 9000
F1-Score. Additionally, YOLOv11 records the lowest average
latency on the GPU at 19.93 ms. In contrast, although
YOLOv10 has the highest Average Precision and slightly
Train Set 6310 faster latency, YOLOv9 is considered the second most
effective overall due to its higher recall and F1-Score, as it
determines the accuracy for real-time detection analysis.
Valid Set 890 Overall, YOLOv11 offers the most balanced and efficient
performance for accident detection in traffic surveillance
systems.

7
B. Accident Detection and Emergency Response Result with minimal false positives and rapid output but sacrifices of
This comparative analysis examines three recent recall, increasing the likelihood of missed events. In contrast,
implementations of YOLO-based traffic accident detection the YOLOv8+Deep SORT system offers the highest overall
systems within smart traffic surveillance frameworks, performance, demonstrating strengths in all critical metrics
emphasizing their algorithmic similarities and comparative such as accuracy, precision, recall, and latency response.
performance. Two of the studies employ YOLOv11 as the
primary object detection architecture, while the third integrates Therefore, based on the data, the YOLOv8+Deep
YOLOv8 with Deep SORT tracking to enhance real-time SORT system outperforms both YOLOv11 models in overall
accuracy and robustness. Although all three aim to address the balance, as it achieved higher scores in accuracy, precision,
need for efficient accident detection, their design priorities and and recall simultaneously. The addition of Deep SORT
optimization strategies differ significantly, resulting in varied enables better object identity tracking, particularly under
performance outcomes. crowded high-traffic scenarios. The recall value of 97.2%
indicates fewer missed detections compared to Arefin’s
The system developed by utilizing YOLOv11 YOLOv11 model, and the F1-score of 95.7% highlights its
demonstrates a well-rounded approach, incorporating vehicle strength in maintaining both high precision and high recall.
damage classification and an integrated emergency alert However, the drawback of the addition of Deep SORT
mechanism [33]. It achieves a detection accuracy of 94.3%, a requires more computational power and storage, which could
precision of 92.8%, a recall of 95.1%, and an F1-score of limit its capability when implemented in a real-world traffic
94.0%. While its inference speed of 0.35 seconds per frame surveillance system. Still, this indicates that the
supports near real-time performance, it is relatively slower YOLOv8+Deep SORT system, through its balanced
compared to the other two systems. The other YOLOv11- optimization of accuracy, precision, and recall, holds greater
based model, on the other hand, prioritizes precision and potential for mitigating traffic-related risks and enhancing
computational efficiency. It records the highest precision at road safety in real-time traffic surveillance applications.
99.40% and the fastest processing speed of 19.93 milliseconds
per frame [15]. However, it reports a significantly lower recall C. AI-Integrated Product Evaluation
value of 82.49%, resulting in a lower F1-score of 90.40%. The comparative analysis between traditional traffic
This reduction in recall suggests a higher risk of missing incident detection and the proposed smart surveillance system
actual accident instances, which may limit the system’s highlights a substantial advancement in road safety response.
suitability in high-risk or safety-critical contexts. The third The systems are compared through different criteria, focusing
system, which integrates YOLOv8 with Deep SORT, delivers more on qualitative data [33]. The study about the comparison
the most balanced and statistically effective performance. It emphasizes the main difference between the two different
results in 98.4% accuracy, 98.5% precision, 97.2% recall, and systems methods of accident prevention, which lies from
a leading F1-score of 95.7% [34]. The inclusion of Deep manual reporting and automated response. The detection
SORT enhances the model's ability to maintain object accuracy from the traditional system is only about 60-70%,
identities in complex traffic scenes, particularly under which is significantly lower than the proposed system having
conditions involving high traffic density. an accuracy of 94.3%. Furthermore, the response time is also
much higher for the former, with its average response time of
10 minutes as opposed to an automated response system of
less than a minute after a car accident.

The proposed smart surveillance system introduces

automation of the entire emergency notification response
system. Unlike traditional existing systems that depend
heavily on human intervention, the smart model utilizes deep
learning to automatically send system notifications to relevant
emergency units. This response system significantly shortens
the delay between accident occurrence and emergency action,
and this eliminates drawbacks associated with human manual
operation. The AI-integrated system also minimizes false
positives and negatives to reduce the likelihood of unnecessary
Figure 3.2.1. Overall Performance Comparison rescue operations and deployment. Moreover, the smart
detection system demonstrates superior scalability and
In comparison to its application, Saravanan’s model environmental adaptability. While traditional systems are
excels in emergency responsiveness and damage typically confined to specific monitored zones, the AI-
categorization but is moderately limited by its processing powered approach can be integrated across extensive road
speed. Arefin’s model is highly efficient in detecting incidents networks. Its ability to function effectively under challenging

8
environmental conditions, such as poor lighting or adverse visibility or adverse weather. These advantages emphasize the
weather, further strengthens its operational reliability. smart system’s effectiveness in enhancing road safety and
Traditional systems often fail in such scenarios, leading to implies that the proposed system is a more efficient, reliable,
missed incidents or delayed responses. and scalable solution compared to the existing traditional
methods.

Criteria Existing System Proposed System IV. CONCLUSION

This study focused on creating a concept for a smart

Accident Detection Manual reporting by Automated real-time traffic surveillance system that can detect real-time road
witnesses or detection using accidents and send emergency alerts immediately. The system
emergency calls YOLO algorithm uses artificial intelligence to monitor live traffic footage and
detect signs of collision faster than manual monitoring. The
system can tell the difference between normal traffic and
Detection Accuracy 60-70% (high false 94.3% (precise actual accidents by using object detection and motion analysis.
positives and identification with
misreporting) deep learning) Through comparison with traditional traffic systems,
the results showed that AI-based systems detect accidents
more accurately and respond much faster. The study found
that models like YOLOv11 and YOLOv8 with Deep SORT
Response Time Delayed due to Instant emergency
gave effective results, especially in identifying accidents
manual reporting (5- alerts (<1 min)
correctly. This system sends automatic alerts with little delay
15 min)
compared to the traditional process that heavily relies on
people and manual input for accident detection and emergency
response alerts.
Emergency Requires manual Automated SMS &
Notification intervention and system alerts to The research also showed that this kind of smart
confirmation responders integrated system works well even in poor weather or low
visibility conditions, which are common problems for regular
road monitoring cameras. This system proves to be more
False Positives / High due to human Low due to AI- useful and reliable in real-world road situations due to the
Negatives error and based classification elements and algorithms that work well with simultaneous
misjudgment accident detection. Although building and installing this type
of system takes time and has a financial counterpart, the
improvements with regards to road safety and emergency
Scalability Limited to specific Can be deployed response will be advantageous.
monitored locations across multiple
smart city networks Conclusively, the use of artificial intelligence in
traffic surveillance provides a more convenient process for
road monitoring, damage prevention, and reduction of fatal
Environmental Ineffective in poor Works efficiently in risks during accidents. The continuous development of smart
Adaptability lighting or bad various conditions monitoring systems contribute to faster transportation by
weather providing faster emergency response and more efficient traffic
monitoring.

Table 3.3.1. Comparison of Existing vs. Proposed System

(Saravanan, 2025)

The smart accident detection system powered by

YOLO-based algorithms clearly outperforms traditional
manual reporting systems in multiple critical aspects. It offers
automated real-time detection, significantly higher accuracy,
faster response times, and reliable emergency notifications
without human intervention. Additionally, it demonstrates low
error rates, superior scalability across the network, and
adaptability under challenging conditions such as low

9
References System: Issues, comprehensive review, and challenges. In
[1] Asian Transport Observatory. (2025). Philippines Road IntechOpen eBooks. [Link]
Safety Profile 2025. [Link]. [15] Arefin, M., Mahin, M., & Mily, F. (2025). Real-time
[Link] rapid accident detection for optimizing road safety in
fetyprofiles/philippines-road-safety-profile-2025/ Bangladesh. Volume 11, Issue 4, e42432.
[2] Thim. (2024, October 29). Road Safety in the Philippines: [Link]
Strategies for Reducing Accidents and Saving Lives - Richest [16] Poongothai, M. (2022). IoT Based Accident Prevention
Philippines. Richest Philippines. [Link] System for Hairpin Bend Roads. Journal of ISMAC The
safety-in-the-philippines-strategies-for-reducing-accidents- Journal of IoT in Social, Mobile, Analytics, and Cloud, 4(3),
and-saving-lives/ 143–152. [Link]
[3] UN Agency. (2021, October 12). The drive to use AI for [17] Selvam, N., Sampath, M. K., Madhavan, J. S., Sharmitha,
safer roads. ITU Hub. [Link] T., & Mugesh, D. S. (2023). IoT based Smart Communication
drive-to-use-ai-for-safer-roads/ System for Accident Prevention. 562–566.
[4] Fernández, J., Cañas, J. M., Fernández, V., & Paniego, S. [Link]
(2021). Robust Real-Time Traffic Surveillance with Deep
Learning. Computational Intelligence and Neuroscience, 2021, [18] Alnashwan, R., Mashaabi, M., Alotaibi, A. A., Qudaih,
1–18. [Link] H., & Al-Braheem, L. (2023). IoT Based Accident Prevention
[5] Khan, T. (2025). AI In Traffic Management. Scribd. System using Machine Learning techniques. 179–188.
[Link] [Link]
management [19] Yusuf, A. A., Asija, D., & Viral, R. (2023). Enhancing
[6] Campbell, O. (2024). The Limitations of Traditional Road Safety: An IoT Based Driver Sleep Detection and
NVR/DVR Systems in Modern Security. Arcadian AI. Alarming System for Accident Prevention.
[Link] [Link]
traditional-nvr-dvr-systems-in-modern-security#:~:text=Lack [20] Mistry, M., Haque, A. M., Mistry, Y., Shah, H., &
%20of%20Advanced%20Video%20Analytics,and Sorathiya, J. (2022). Iot based railway accident prevention and
%20analysis%20of%20potential%20threats monitoring system. International Journal of Engineering
[7] NSoft Vision. (2025). Traditional vs. AI Video Applied Sciences and Technology, 7(5), 207–213.
Surveillance: A Comparative Insight. [Link]
[Link] [21] Ye, Q., & Chen, R. (2019). Vehicle collision detection
[8] Dilek, E., & Dener, M. (2023). Computer Vision method based on machine learning.
Applications in Intelligent Transportation Systems: A survey. [22] Zhang, Y., Zha, W., Wang, C., Zhao, S., Pan, X., &
Sensors, 23(6), 2938. [Link] Wang, W. (2020). Machine learning method and device for
intelligent data collision.
[9] Mandal, V., Mussah, A. R., Jin, P., & Adu-Gyamfi, Y.
(2020). Artificial Intelligence-Enabled Traffic Monitoring [23] Elsagheer, S. A., Atallah, Y., & Hashem, H. (2023).
System. Sustainability, 12. [Link] Enhancing Road Safety Using the Internet of Vehicles: A
[10] Skrabacz, A., Bsoul-Kopowska, M., & Kozicki, B. Machine Learning-Based Collision Detection Approach. 68–
(2024). The Application of Artificial Intelligence in Road 72. [Link]
Traffic Management and Its Safety Improvement. Transport [24] Wang, S., Sun, J., Wang, R., Dai, X., & Meng, Y. (2017).
Problems, 19(4). [Link] Vehicle targets recognizing method and anti-collision device
[11] Pathik, N., Gupta, R. K., Sahu, Y., Sharma, A., Masud, using multi-line point cloud data machine learning.
M., & Baz, M. (2022). AI Enabled Accident Detection and [25] Khatun, M., Jung, R., & Glaß, M. (2023). Scenario-based
Alert System Using IoT and Deep Learning for Smart Cities. collision detection using machine learning for highly
Sustainability, 14(13), 7701. automated driving systems. Systems Science & Control
[Link] Engineering, 11(1).
[12] Muddam, Y. (2024). Crash Detecting System Using Deep [Link]
Learning. Electronic Theses, Projects, and Dissertations. [26] Ahmed, M. I. B., Zaghdoud, R., Ahmed, M. S., Sendi, R.,
[Link] Alsharif, S., Alabdulkarim, J., Albin Saad, B. A., Alsabt, R.,
article=3138&context=etd Rahman, A., & Krishnasamy, G. (2023). A Real-Time
[13] Kabir, M. F., & Roy, S. (2022). Real-time vehicular Computer Vision Based Approach to Detection and
accident prevention system using deep learning architecture. Classification of Traffic Incidents. Big Data and Cognitive
Expert Systems with Applications, 206, 117837. Computing, 7(1), 22. [Link]
[Link] [27] Atzori, A., Barra, S., Carta, S., Fenu, G., & Podda, A. S.
[14] Gawande, U., Hajari, K., & Golhar, Y. (2020). (2021, March 1). Heimdall: an AI-based infrastructure for
Pedestrian Detection and Tracking in Video Surveillance traffic monitoring and anomalies detection. IEEE Xplore.

10
[Link]
52
[28] Al-Agroudy, Z., Mohamed, A., Ashraf, Z., Al-Sayed, S.,
Gad, W., Hassan, Z., & Reda, R. (2023). AI-Safe
Transportation: Real-Time Incident Detection and Alerting
System in Smart Cities. 2023 Eleventh International
Conference on Intelligent Computing and Information
Systems (ICICIS), 523–529.
[Link]
[29] Bajwa, A. (2025). AI-based Emergency Response
Systems: A Systematic Literature Review on Smart
Infrastructure Safety. SSRN Electronic Journal.
[Link]
[30] Tahir, M., Qiao, Y., Kanwal, N., Lee, B., & Ashgar, M.
(2023, December). Real-Time Event-Driven Road Traffic
Monitoring System Using CCTV Video Analytics | IEEE
Journals & Magazine | IEEE Xplore. [Link].
[Link]
[31] Alvi, U., Khattak, M. A. K., Shabir, B., Malik, A. W., &
Muhammad, S. R. (2020). A Comprehensive Study on IoT
Based Accident Detection Systems for Smart Vehicles. IEEE
Access, 8, 122480–122497.
[Link]
[32] Satheesh, N., Gopisankar, N., Kumarganesh, S.,
Anthoniraj, S., Saravanakumar, S., Sagayam, K. M., Pandey,
B. K., & Pandey, D. (2025). Advanced AI-driven emergency
response systems for enhanced vehicle and human safety. Iran
Journal of Computer Science. [Link]
025-00228-w
[33] Saravanan, P., Karthick, B., & Kalaiarasu, M. (2025).
Smart Traffic Accident Detection and Automated Emergency
Response System Using Object Detection Algorithm. IJRTI,
10(4). [Link]
[34] Karim, A., Raza, M. A., Alharthi, Y. Z., Abbas, G.,
Othmen, S., Hossain, M. S., Nahar, A., & Mercorelli, P.
(2024). Visual Detection of Traffic Incident through
Automatic Monitoring of Vehicle Activities. World Electric
Vehicle Journal, 15(9), 382.
[Link]

11