VISVESVARAYA TECHNOLOGICAL UNIVERSITY

Jnana Sangama, Belagavi – 590 018.

Technical Seminar Report

On

“A novel approach to predict the traffic accident

assistance based on deep learning”
Submitted in partial fulfilment of the requirement for the curriculum of the 8th Semester

Bachelor of Engineering
In
Artificial Intelligence and Machine Learning
Submitted by
N PUNITH KUMAR : 1VI21AI029

Under the supervision of

DR. KANTHARAJU H C

Associate Professor & HoD

DEPARTMENT OF ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

VEMANA INSTITUTE OF TECHNOLOGY
BENGALURU – 560034
2024-2025
 Karnataka Reddy Jana Sangha®
VEMANA INSTITUTE OF TECHNOLOGY
(Affiliated to Visvesvaraya Technological University, Belagavi)
Koramangala, Bengaluru-560034.

DEPARTMENT OF
ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

CERTIFICATE
This is to certify that Technical Seminar work entitled “A novel approach to predict the
traffic accident assistance based on deep learning” carried out by N PUNITH KUMAR
(1VI21AI029) is a bonafide student of Vemana Institute of Technology in partial
fulfilment of the award of Bachelor of Engineering in Artificial Intelligence and
Machine Learning of the Visvesvaraya Technological University, Belagavi during
the academic year 2024-25. It is certified that all suggestions indicated for internal
assessment have been incorporated in the report. The Technical Seminar report has
been approved as it satisfies the academic requirements in respect of the Seminar work
prescribed for the said degree.

Guide Associate Professor & HoD Principal

(Ms. Lakshmi Sravani G) (Dr. Kantharaju H C) (Dr. Vijayasimha Reddy B.G)
 ACKNOWLEDGEMENT
I sincerely thank Visvesvaraya Technological University for providing a
platform to do the Technical Seminar.

Firstly, I would like to express my deep sense of gratitude to our institute

“Vemana Institute of Technology” that provided us an opportunity to do a
Technical Seminar entitled “A novel approach to predict the traffic accident
assistance based on deep learning”.

I thank Dr. Vijayasimha Reddy. B. G, Principal, Vemana Institute of

Technology, Bengaluru for providing the necessary support.

I would like to place on record our regards to Dr. Kantharaju H C, Associate

Professor and Head of Department of Artificial Intelligence and Machine
Learning for his continued support.

I would like to thank our Technical Seminar coordinators, Mr. Harish Kumar
M , Assistant Professor, Dept. of AI&ML for their support and coordination.

I would like to thank our Technical Seminar guide Ms. Lakshmi Sravani
G, Assistant Professor, Dept. of AI&ML for his continuous support
and valuable guidance towards successful completion of the Technical
Seminar.

I would be failing in my duty if myself do not thank the faculty members,

batch mates, lab staffs, technicians and family members for their constant support
and guidance.

Date: 21/4/2025 N PUNITH KUMAR

Place: Bengaluru (1VI21AI029)

i
 ABSTRACT
Road traffic accidents are a major global concern, claiming thousands of lives
each year. Predicting the necessity of medical assistance in such incidents is
crucial for efficient emergency response and resource allocation. This
technical seminar report presents a novel deep learning-based framework that
aims to predict whether a traffic accident requires medical assistance. The
proposed system is designed to be generalizable and adaptable across various
cities and datasets. The architecture comprises three key stages: a pre-
processing stage that handles data cleaning, transformation, and resampling; a
post-processing stage that applies genetic algorithms and boosting techniques
to assign feature importance; and a final model training stage that utilizes a
customized two-dimensional Convolutional Neural Network (2D-CNN) to
perform the actual prediction. The model converts tabular accident data into
matrices to effectively train CNN. Experimental evaluations were conducted
using datasets from six different UK cities, and the model’s performance was
benchmarked against six state-of-the-art machine learning classifiers. The
results demonstrated that the proposed model achieved superior accuracy and
F1-scores in all cities, confirming its robustness, real-time capability, and
potential application in intelligent transportation systems.

ii
 TABLE OF CONTENTS
Chapter No. Title Page No.

Acknowledgement i

Abstract ii

Table of contents iii-iv

List of Figures v

List of Abbreviations vi

1 INTRODUCTION 1

2 LITERATURE SURVEY 2

3 DESIGN METHODOLOGY 3

3.1 Pre-processing Stage 3-4

3.2 Post-processing Stage 4

3.3 CNN-based Prediction Model 4-5

4 IMPLEMENTATION 6

4.1 Dataset and Feature Classification 6

4.2 Data Filtering and Oversampling 6

4.3 Model Training and Optimization 7

4.4 Validation and Evaluation 7

5 RESULTS 8-9

CONCLUSION 10

iii
 TABLE OF CONTENTS
Chapter No. Title Page No.

FUTURE ENHANCEMENTS 11

REFERENCES 12

BIBLIOGRAPHY 13

iv
 LIST OF FIGURES
Figure No Title Page No

Figure 3.1 Example of samples before and after cleaning process 3

Figure 3.2 Architecture summary of CNN-2D model 5

F1-score of CNN-2D in training and test set for South

Figure 5.1 wark 8

Figure 5.2 F1-Score by city for Assistance 9

v
 LIST OF ABBREVIATIONS

ANPR : Automatic Number Plate Recognition

CPU : Central Processing Unit

CUDA : Compute Unified Device Architecture

DNN : Deep Neural Network

DIP : Digital Image Processing

FPS : Frames Per Second

GPU : Graphics Processing Unit

HCI : Human–Computer Interaction

MAE : Mean Absolute Error

MSE : Mean Squared Error

RMSE : Root Mean Squared Error

R² : Coefficient of Determination

SAE : Stacked Autoencoder

VGG : Visual Geometry Group

VIP : Video Image Processing

CHAPTER 1
INTRODUCTION

With the rapid increase in global mobility, road traffic accidents have become a major concern
for governments and urban authorities. Every day, millions of vehicles operate in dynamic
environments, often leading to unpredictable incidents, including accidents that range from minor
damage to life-threatening injuries. Thousands of lives are lost annually due to road accidents,
underlining the need for intelligent and proactive intervention systems.
In recent years, Artificial Intelligence (AI) has been widely adopted across various fields,
including healthcare, manufacturing, and transportation. In the context of traffic safety, AI-driven
solutions offer the potential to not only analyze and understand traffic patterns but also predict
and respond to accidents more effectively. Among the key challenges is the prediction of
medical assistance following a traffic accident — a task that requires advanced data analysis,
real-time decision-making, and generalizable modeling across diverse environments.
Traditionally, researchers have relied on statistical models to assess the severity of accidents.
manufacturing, and transportation. In the context of traffic safety, AI-driven solutions offer the
potential to not only analyze and understand traffic patterns but also predict and respond to
accidents more effectively. To overcome these limitations, machine learning (ML) and deep
learning (DL) techniques have gained prominence.
This report focuses on a novel framework that uses a two-dimensional Convolutional Neural
Network (2D-CNN) to predict whether a traffic accident will require medical assistance. The
framework incorporates a multi-stage pipeline:
1. A pre-processing stage to clean, transform, filter, and balance the dataset.
2. A post-processing stage that uses genetic algorithms to optimize feature weights.
3. A final model training stage where the weighted data is transformed into matrices and
processed by the CNN.
To evaluate its generalization capabilities, the model was tested on traffic datasets from six
different cities in the United Kingdom. It demonstrated significantly higher accuracy and
robustness compared to other state-of-the-art ML models, including Naive Bayes, Support Vector
Machines, and Multilayer Perceptrons.

Dept. of AI&ML, Vemana IT Page-1 2024-25

CHAPTER-02
LITERATURE SURVEY

The prediction and classification of traffic accident severity have been extensively studied over
the past decades, with researchers employing a wide range of models and techniques. These
models typically rely on assumptions regarding the distribution of input data, such as normality
or linearity. While these approaches offer interpretability, they suffer from limited flexibility and
reduced performance when applied to real-world, heterogeneous datasets that contain missing
values, outliers, or imbalanced class distributions. To address these limitations, machine learning
(ML) and evolutionary computing techniques have gained traction. Techniques such as genetic
algorithms, fuzzy logic, and neural networks have been successfully applied to optimize model
parameters, discover hidden patterns, and classify accident severity. For example, several studies
have used genetic algorithms to enhance feature selection and improve the performance of
classifiers in accident prediction tasks.
These methods have been shown to outperform traditional approaches, particularly when
dealing with large and complex datasets.Further improvements have been achieved by
integrating deep learning models, especially Convolutional Neural Networks (CNNs). Recent
research has focused on developing CNN-based architectures to predict accident outcomes,
including whether a medical response is needed. These models have shown promising results,
particularly in classifying accidents based on their severity levels—such as distinguishing
between property damage-only cases and those involving injuries or fatalities. Despite these
developments, many existing models remain domain-specific, performing well only on datasets
from specific regions.
Techniques such as genetic algorithms, fuzzy logic, and neural networks have been successfully
applied to optimize model parameters, discover hidden patterns, and classify accident severity
Techniques such as genetic algorithms, fuzzy logic, and neural networks have been successfully
applied to optimize model parameters, discover hidden patterns, and classify accident severity
The lack of generalizability makes it difficult to deploy such models across multiple cities or
countries. The current work addresses this gap by proposing a generalizable framework that can
be applied to various urban environments with minimal reconfiguration, ensuring adaptability
and consistency in predicting accident severity and required assistance.

Dept. of AI&ML, Vemana IT Page-2 2024-25

CHAPTER-03
DESIGN METHODOLOGY
The proposed model for predicting medical assistance in traffic accidents is designed as a
generalizable three-stage architecture that can be applied to any city or dataset with minimal
changes. These stages include:
1. Pre-processing Stage
2. Post-processing Stage
3. CNN-based Prediction Model
The key goal is to build a model that transforms raw accident records into a structured input for a
Convolutional Neural Network (CNN), enabling it to classify the severity of accidents and
determine the need for medical intervention.

3.1 RESEARCH METHODOLOGY

The first stage involves preparing the raw dataset for effective learning. This includes a sequence
of four critical operations:

1. Cleaning

Figure 3.1 Example of samples before and after cleaning process

Redundant, missing, null, and outlier records are removed to ensure uniformity in the dataset.

Dept. of AI&ML, Vemana IT Page-3 2024-25

2. Transformation:
Selected variables are categorized into five meaningful groups for prediction:
1. Location and Scale (e.g., accident coordinates, number of vehicles)
2. Driving Limitations (e.g., surface condition, speed limit)
3. Temporal and Environmental (e.g., lighting, weather, weekday)
4. Vehicle (e.g., vehicle type, age, point of impact)
3. Filtering Areas:
City regions are divided into rectangles based on accident coordinates. Only areas with both
slight and assistance accidents are kept — balancing the dataset while preserving spatial
information.
4. Resampling:
Despite filtering, class imbalance remains. City regions are divided into rectangles based on
accident coordinates. To counter this, Borderline SMOTE-II is applied on training data to
synthetically generate minority samples (assistance-needed cases).

3.2 PERFORMANCE EVALUATION CRITERIA

Once the dataset is clean and balanced, a critical transformation is applied to prepare the data for
CNN input.
1. Feature Importance via Genetic Algorithm:
A boosting algorithm is used to assign weights to each feature based on its importance. These
weights are optimized using a Genetic Algorithm, which evolves hyperparameters (e.g., learning
rate, depth) over generations to maximize the F1-score.
2. Matrix Construction :
The weighted features are converted into an n×n matrix where important features are placed in
central positions and less important ones near the edges. This mimics how CNNs process images
by focusing more on the center.

3.3 CNN-based Prediction Model

The final stage involves a 2D Convolutional Neural Network (CNN-2D) with four
convolutional layers and one fully connected (dense) layer:

Dept. of AI&ML, Vemana IT Page-4 2024-25

 Figure 3.2 – Architecture summary of CNN-2D model

A convolutional neural network (CNN), is a network architecture for deep learning which learns
directly from data. CNNs are particularly useful for finding patterns in images to recognize
objects. They can also be quite effective for classifying non-image data such as audio, time series,
and signal data.
The architecture consists of these layers:
1. Conv Layers: 64, 512, 128, 256 filters respectively, with 3×3 kernel size, ReLU
activation, and batch normalization.
2. Flattening Layer: Converts the feature map into a 1D vector.
3. Dense Layer: Includes 256 nodes connected to the output layer.
4. Output Layer: Uses the Softmax activation function to classify accident cases as either:
1. Medical assistance required
2. No assistance required

The model is optimized using binary cross-entropy loss and trained to ensure real-time
inference with high accuracy and minimal overfitting. CNNs are particularly useful for finding
patterns in images to recognize objects. The final stage involves a 2D Convolutional Neural
Network (CNN-2D) with four convolutional layers and one fully connected (dense) layer.

Dept. of AI&ML, Vemana IT Page-5 2024-25

CHAPTER 4
IMPLEMENTATION
The implementation of the proposed deep learning model involves the preparation of the dataset,
transformation of features, and application of genetic algorithms for model optimization. This
chapter provides detailed insights into the dataset characteristics, data handling techniques, and
training pipeline for the Convolutional Neural Network (CNN-2D) used for predicting traffic
accidents that require medical assistance.

4.1 Dataset and Feature Classification

The dataset used in this work was obtained from the UK Government’s Department for
Transport, covering road traffic accidents from six cities: Southwark, Manchester, Birmingham,
Liverpool, Sheffield, and Cornwall. These datasets include accident records from 2005 to 2019
and contain up to 81 features describing the accident circumstances.To make the dataset
compatible with the model and generalizable across cities, a strict feature classification and
filtering process was adopted highly correlated features (correlation > 0.5) were eliminated to
reduce redundancy.
Features were grouped under five predefined categories:
1. Location & Scale
2. Driving Limitations
3. Temporal & Environmental
4. Vehicle-related
5. Victim-related
Following this process, 17 representative features were retained for training, ensuring a balance
between comprehensiveness and model efficiency.

4.2 Data Filtering and Oversampling

Traffic accident datasets typically suffer from class imbalance—with far more records of minor
accidents than those requiring assistance. To handle this, the following steps were implemented:
Area Filtering: Cities were divided into rectangular spatial grids based on coordinates. This
helped balance the dataset while preserving geographical diversity.

Dept. of AI&ML, Vemana IT Page-6 2024-25

4.3 Model Training and Optimization
Once the dataset was prepared, model training was conducted using a 2D Convolutional Neural
Network optimized through genetic algorithms. The training process followed these key steps:
1. Matrix Construction: Each accident record was converted into a matrix using the weights
assigned to its features. These weights were determined by a boosting model optimized
using a genetic algorithm.
2. Hyperparameter Optimization: The genetic algorithm optimized key hyperparameters
such as:
1. ETA (learning rate)
2. Maximum tree depth
3. Minimum child weight
3. The fitness function for optimization was the F1-score of the boosting model applied on
the validation dataset.
4. The final optimized weights were used to generate input matrices for the CNN.

4.4 Validation and Evaluation

The trained CNN-2D model was validated using unseen test data for each city. The results were
compared against six state-of-the-art models:
1. Naive Bayes (NB)
2. Support Vector Classifier (SVC)
3. K-Nearest Neighbors (KNN)
4. Random Forest (RF)
5. Logistic Regression (LR)
Evaluation metrics included:
1. Precision
2. Recall
3. F1-Score
4. Binary Cross-Entropy
In all cities, the proposed CNN-2D model consistently outperformed others in terms of F1-score,
particularly for predicting assistance-required accidents in imbalanced or complex scenarios.

Dept. of AI&ML, Vemana IT Page-7 2024-25

CHAPTER 5
RESULTS

The proposed CNN-2D model was evaluated using traffic accident datasets collected from six
different UK cities: Southwark, Manchester, Birmingham, Liverpool, Sheffield, and Cornwall.
The goal was to assess how well the model could generalize across varying regional conditions
while maintaining high predictive performance, especially in identifying accident cases requiring
medical assistance.

Figure 5.1: F1-score of the CNN-2D in training and test set for Southwark

During testing, the CNN-2D model consistently outperformed traditional machine learning
models highest accuracy and F1-score were achieved for the Southwark dataset such as Naive
Bayes (NB), Support Vector Classifier (SVC), K-Nearest Neighbors (KNN), Random Forest
(RF), Polynomial Logistic Regression (LR), and Multi-Layer Perceptron (MLP). Among all
cities, the highest accuracy and F1-score were achieved for the Southwark dataset, followed
closely by Manchester and Birmingham. Performance was measured using several key metrics
such as precision, recall, F1-score, Mean Absolute Error (MAE), Mean Squared Error (MSE),
Root Mean Squared Error (RMSE), and Coefficient of Determination (R²).

Dept. of AI&ML, Vemana IT Page-8 2024-25

 Figure 5.2: F1-Score by city for Assistance

The CNN-2D model achieved the lowest MAE and MSE across all cities, indicating a strong
ability to reduce prediction error. In terms of classification quality, it maintained a high F1-score
even on test datasets with complex or imbalanced class distributions.
Comparatively, the CNN-2D model achieved an F1-score of 0.91 in Southwark, whereas the best
competing model, MLP, scored 0.84. Similar trends were observed across Manchester (CNN:
0.90, SVC: 0.81), Birmingham (CNN: 0.88, KNN: 0.80), and the remaining cities. During testing,
the CNN-2D model consistently outperformed traditional machine learning models highest
accuracy and F1-score were achieved for the Southwark dataset such as Naive Bayes (NB),
Support Vector Classifier (SVC), K-Nearest Neighbors (KNN), Random Forest (RF), Polynomial
Logistic Regression (LR), and Multi-Layer Perceptron (MLP). Among all cities, the highest
accuracy and F1-score were achieved for the Southwark dataset, followed closely by Manchester
and Birmingham. These results validate the model’s effectiveness and its potential for deployment
in real-world traffic systems.

Dept. of AI&ML, Vemana IT Page-9 2024-25

 CONCLUSION

This work presented a novel and generalizable deep learning-based framework for predicting
whether a traffic accident will require medical assistance. The proposed system combined a
structured three-stage approach involving data pre-processing, post-processing using genetic
algorithms, and classification through CNN-2D architecture.
By converting raw tabular traffic accident data into structured input matrices and optimizing
feature weights, the model demonstrated the capability to learn meaningful patterns without relying
on domain-specific assumptions.
The framework proved robust across diverse datasets from six different UK cities, outperforming
conventional machine learning models in terms of F1-score, accuracy, and error metrics such as
MAE and MSE. The proposed system combined a structured three-stage approach involving data
pre-processing, post-processing using genetic algorithms, and classification through CNN-2D
architecture.
Moreover, the use of Borderline SMOTE-II for class balancing and the integration of genetic
algorithm-driven feature weighting significantly enhanced model sensitivity to minority class
cases, accidents requiring medical attention. The model maintained a high degree of consistency
across both densely populated urban regions and less populated rural areas.
In conclusion, the proposed method not only improves the prediction of accident severity but also
lays the groundwork for developing intelligent, real-time systems that can assist emergency
response units in prioritizing medical aid, ultimately contributing to more efficient and life-saving
transportation systems.

Dept. of AI&ML, Vemana IT Page-10 2024-25

 FUTURE ENHANCEMENTS

While the proposed CNN-2D-based framework has shown promising results in predicting the
need for medical assistance in traffic accidents, several enhancements can be incorporated to
improve its practical deployment, scalability, and overall intelligence.
1. Integration of Real-Time Data Sources Future systems can incorporate live traffic
feeds, weather updates, and road condition sensors to dynamically update accident risk
and severity predictions. This would make the model suitable for real-time emergency
dispatching.
2. Automatic Number Plate Recognition (ANPR) Integrating ANPR systems would allow
authorities to automatically identify and notify vehicle owners involved in severe
accidents. This would also support automated documentation and traffic violation
tracking.
3. Multi-Modal Data Fusion The inclusion of audio data (e.g., crash sounds), textual
reports (e.g., emergency call transcripts), or even drone footage could help the system
make more comprehensive assessments.
4. Cloud-Based Deployment and Scalability Hosting the prediction engine on cloud
platforms could support continuous learning, easy access for traffic control centers, and
simultaneous deployment across multiple regions.
5. Extension to Emergency Resource Allocation The framework can be expanded to not
only predict the severity of accidents but also recommend the type and number of
emergency resources (ambulances, rescue units) required based on predicted severity and
nearby availability.
6. Transfer Learning and Federated Learning To support privacy and adaptability across
countries or organizations, federated learning can be employed, allowing local models to
improve globally without sharing raw data.
By addressing these future directions, the system can evolve into a fully integrated, intelligent,
and life-saving component of next-generation smart traffic management and emergency response
networks.

Dept. of AI&ML, Vemana IT Page-11 2024-25

 REFERENCES
[1]. Liu Shuo, Gu Yuhai, Rao Wenjun, Wang Juyuan, “Traffic Violation Detection Using
YOLOv3 in Intersection Images,” 2021 IEEE International Conference on CVPR.
[2]. Maha Vishnu V. C., Rajalakshmi M., Nedunchezhian R., “Dynamic Traffic Signal
Control and Violation Detection,” International Journal of Intelligent Systems and
Applications, 2018.
[3]. BIT-Vehicle Dataset – Available online: https://github.com/PKU-IMRE/Bit-vehicle
(Accessed 2025).
[4]. ApolloScape Dataset – Available online: http://apolloscape.auto/scene.html (Accessed
2025).
[5]. Kalman R. E., “A New Approach to Linear Filtering and Prediction Problems,”
Transactions of the ASME–Journal of Basic Engineering, 1960.
[6]. OpenCV Library Documentation – https://docs.opencv.org/
[7]. CUDA Toolkit Documentation – https://developer.nvidia.com/cuda-zone
[8]. Zhang, Z., “Histogram Equalization Techniques for Image Enhancement,” International
Journal of Computer Applications, 2015.
[9]. Redmon, J., Farhadi, A., “YOLOv3: An Incremental Improvement,” arXiv preprint
arXiv:1804.02767, 2018.
[10]. José F. Vicent, Juan M. Alberola, Begoña Pérez, and Manuel Pérez-Rodríguez,
“A novel approach to predict the traffic accident assistance based on deep learning,”
Neural Computing and Applications, vol. 36, pp. 5345–5368, 2024.

Dept. of AI&ML, Vemana IT Page-12 2024-25

 BIBLIOGRAPHY

[1]. Goodfellow, Ian, Bengio, Yoshua, and Courville, Aaron,

Deep Learning, MIT Press, 2016.
[2]. Géron, Aurélien,
Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, O'Reilly Media,
2022.
[3]. Aggarwal, Charu C.,
Data Classification: Algorithms and Applications, CRC Press, 2014.
[4]. Brownlee, Jason,
Imbalanced Classification with Python, Machine Learning Mastery, 2020.
[5]. Smola, A. J., and Schölkopf, B.,
“A Tutorial on Support Vector Regression,” Statistics and Computing, vol. 14, no. 3, pp.
199– 222, 2004.
[6]. He, Haibo, and Garcia, Edwardo A.
“Learning from Imbalanced Data,” IEEE Transactions on Knowledge and Data
Engineering, vol. 21, no. 9, pp. 1263–1284, 2009.
[7]. Chawla, Nitesh V., et al.,
“SMOTE: Synthetic Minority Over-sampling Technique,” Journal of Artificial
Intelligence Research, vol. 16, pp. 321–357, 2002.

Dept. of AI&ML, Vemana IT Page-13 2024-25