Mid Term Report

On
Distracted Driver Detection
Project-I

BACHELOR OF TECHNOLOGY
(Artificial Intelligence and Data Science)

SUBMITTED BY:
Manya Prakash
2231159
Sounak Bhowal
2231181
Vanshika Sharma
2231185
Date- 9th April 2026

Under the Guidance of

Dr. Yogesh Shahare
Assistant Professor

Department of Artificial Intelligence and Data Science

Chandigarh Engineering College
Jhanjeri, Mohali-140307
1
 Table of Contents

S.No. Contents Page No

1. Introduction 3-6

2. System Requirements 7-8

3. Software Requirement Analysis 9-10

4. Software Design 11

5. Implementation 12-14

6. References 15-16

2
 Chapter-1
Introduction
Distracted driving has become a major public safety concern, contributing significantly to road
accidents, injuries, and fatalities worldwide. Any activity that diverts a driver's focus from operating a
vehicle, such as texting, talking on the phone, adjusting in-vehicle systems, eating, or interacting with
passengers, falls under distracted driving. The rapid integration of technology into daily life has further
exacerbated this issue, making it imperative to develop advanced detection and prevention systems.

1.1 The Scope and Impact of Distracted Driving

Distracted driving is a critical public safety issue, contributing significantly to road accidents
globally. A 2024 survey by The Zebra in the U.S. revealed that 47% of drivers admitted to texting
while driving, 60.2% talked on the phone, and 57% adjusted GPS devices. A cross-sectional survey in
New South Wales and Western Australia (1,347 drivers, ages 18–65) identified prevalent distractions:
loss of concentration (71.8%), adjusting in-vehicle equipment (68.7%), observing external stimuli
(57.8%), and passenger conversations (39.8%). On average, drivers engage in distracting activities
approximately once every six minutes, markedly increasing accident risk. In the U.S., the National
Highway Traffic Safety Administration (NHTSA) reported that in 2022, 8% of fatal crashes (3,308
deaths) and 13% of police-reported crashes were distraction-related, resulting in 289,310 injuries.

Regional variations highlight the global scope of the issue. For example, Germany reported 8,233
injuries and 117 deaths in distraction-related crashes in 2021, though newer data is needed. Younger
drivers (ages 18–30) are particularly vulnerable, with The Zebra’s 2024 survey noting that 55% of Gen
Z and Millennials texted while driving, compared to 33% of Boomers. These drivers often perceive
distractions as less hazardous, contributing to their higher involvement in distraction-related accidents.
The frequent engagement in distracting activities underscores the urgent need for targeted interventions
and public awareness campaigns to mitigate risks.

1.2 Effects of Distraction on Driving Performance

Distracted driving severely impairs critical driving functions, including reaction time, lane-
keeping, and situational awareness. Texting while driving increases accident risk by 400%, as it diverts
visual, manual, and cognitive attention from the road. Other distractions, such as using an onboard
computer (44% increased risk), misusing driver-assist functions like lane assist (56% increased risk),
and adjusting car radio settings via onboard systems (89% increased risk), also significantly elevate
crash likelihood, though these specific percentages require study-specific verification. Naturalistic
driving studies indicate that drivers spend approximately 14.5% of their driving time engaged in
3
distractions, with Cambridge Mobile Telematics (CMT) reporting phone interactions in 58% of trips in
2022.

The severity of distractions is evident in their impact on driving behavior. Sending or reading a text
takes a driver’s eyes off the road for about 5 seconds, equivalent to traveling the length of a football
field blind at 55 mph. Reaching for an object can increase crash risk by 800%, highlighting the dangers
of even brief distractions. The high prevalence of phone use during driving, as reported by CMT,
correlates directly with increased collision rates, emphasizing the need for technologies and policies to
reduce driver inattention and enhance road safety.

1.3 Challenges in Enforcement and Regulation

Enforcing distracted driving laws is challenging due to lower social stigma compared to
offenses like drunk driving and difficulties in proving distraction at crash scenes. In the U.S., 48 states
have banned texting while driving, and 24 prohibit handheld phone use, but enforcement and
compliance remain inconsistent. Germany introduced “distraction” as a formal category in police
accident reports in 2021, documenting 8,233 injuries and 117 deaths in distraction-related crashes,
though more recent data is unavailable. Underreporting is a significant issue, with the NHTSA
estimating that 29% of U.S. traffic deaths and injuries (approximately 12,400 deaths in 2021) may be
distraction-related, far exceeding the reported 8% of fatal crashes (3,308 deaths in 2022).

The lack of uniform reporting standards across jurisdictions complicates efforts to address distracted
driving. In the U.S., inconsistent state-level data collection leads to under- or over-reporting of
distraction-related incidents. Unlike drunk driving, which has clear testing protocols, proving
distraction often relies on subjective evidence, such as witness accounts or driver admissions. Public
perception also plays a role, as many drivers do not view distractions like phone use as inherently
dangerous, reducing the deterrent effect of existing laws and necessitating stronger enforcement
strategies and educational initiatives.

1.4 Young Drivers and the Psychology of Distraction

Young drivers (ages 18–24) are disproportionately affected by distracted driving, with 55% of
Gen Z and Millennials admitting to texting while driving, compared to 47% of all drivers, according to
The Zebra’s 2024 survey. Approximately 40% of young drivers reported sending or reading messages,
reflecting a 31% increase in texting while driving from 2021–2024, consistent with a reported 2.5-fold
rise from 2016–2022. Psychological studies indicate that young drivers overestimate their multitasking
abilities, underestimating the risks of cognitive overload, which contributes to their higher involvement
in distraction-related crashes.
4
The Centers for Disease Control and Prevention (CDC) notes that drivers aged 15–20 are the most
distracted, with 7% of fatal crash-involved drivers in this group reported as distracted, the highest
proportion among age groups. This overconfidence in multitasking is compounded by the widespread
use of smartphones, which are integral to young drivers’ daily lives. The significant increase in texting
behavior over recent years highlights the need for targeted interventions, such as educational
campaigns and technology-based solutions, to address the unique psychological and behavioral factors
driving distraction among younger populations.

1.5 The Role of Technology in Distracted Driving Detection

Modern vehicles integrate advanced infotainment systems, voice-controlled assistants, and
driver-assist features for convenience, but these can introduce new sources of distraction. Conversely,
AI-powered driver monitoring systems (DMS) using infrared cameras, gaze-tracking algorithms, and
real-time alerts offer promising solutions to mitigate these risks. The National Distracted Driving
Coalition (NDDC) reports that vehicles equipped with DMS have fewer insurance claims,
demonstrating their effectiveness in reducing distraction-related incidents.

The dual role of technology underscores the need for balanced implementation. While in-vehicle
technologies like touchscreens can divert attention, DMS technologies counteract these risks by
actively monitoring driver behavior and issuing alerts when inattentiveness is detected. The NHTSA
highlights the contribution of complex infotainment systems to distraction, but advancements in AI-
driven monitoring provide a counterbalance, enabling real-time interventions that prevent accidents and
promote safer driving habits.

1.6 Deep Learning-Based Approaches to Detecting Distracted Driving

Machine learning (ML) and deep learning (DL) models are increasingly utilized to combat
distracted driving by analyzing driver behavior, eye movements, and hand positions. These AI-powered
models can classify distractions with high accuracy, often exceeding 90% in controlled studies.
Convolutional Neural Networks (CNNs) are particularly effective in image-based driver monitoring
systems, distinguishing safe driving from distractions such as texting, drinking, or reaching for objects.
The State Farm Distracted Driver Detection dataset serves as a robust foundation for training these
models, with transfer learning using architectures like ResNet, MobileNetV2, and EfficientNet
enhancing efficiency and accuracy.

The practical deployment of these models is facilitated by frameworks like OpenCV and Flask,
enabling real-time implementation in smart vehicles and road safety applications. The high accuracy of
CNNs, when trained on comprehensive datasets like State Farm, supports their integration into driver
monitoring systems, offering a proactive approach to accident prevention. Ongoing research continues
5
to refine these models, improving their adaptability to diverse driving conditions and further
strengthening their role in addressing the global challenge of distracted driving.

6
 Chapter-2
System Requirements
This section outlines the minimum and recommended hardware and software configurations needed to
run the project smoothly. The system should be capable of handling image processing tasks without
performance issues.

2.1 Hardware Requirements

2.1.1 Processor
A minimum of Intel Core i3 or AMD equivalent is required. For faster processing, an Intel Core
i5/i7 or higher is recommended.

2.1.2 RAM
At least 4 GB of RAM is necessary. However, 8 GB or more is recommended for smooth
execution, especially when handling large image datasets.

2.1.3 Storage
A minimum of 2 GB of free disk space is required to store datasets and output results. SSD storage
is preferred for faster file access.

2.1.4 Display
A standard HD display is sufficient. A larger screen with higher resolution helps while working
with visualization tasks.

2.2 Software Requirements

2.2.1 Operating System
Compatible with Windows 10/11, Linux (Ubuntu), or macOS. Any OS supporting Python 3.x can
run the project.

2.2.2 Programming Language

Python 3.7 or above should be installed. The code is optimized for modern Python environments.

2.2.3 Python Libraries

The following Python libraries are required:
OpenCV: for image processing
NumPy: for numerical computations
Matplotlib: for image visualization
7
os, zipfile: for file and directory management

2.2.4 Development Environment

Jupyter Notebook is preferred for ease of visualization and step-by-step execution. Alternatively,
any IDE like VS Code or PyCharm can be used.

8
 Chapter-3
Software Requirement Analysis
Software requirement analysis focuses on identifying the necessary tools, libraries, and resources
needed to successfully develop and execute the image processing project. This phase ensures that both
functional and non-functional needs are clearly understood and addressed before implementation
begins.

3.1 Purpose of the Project

The main objective of this project is to process and visualize images using Python and
OpenCV. It involves reading images, applying preprocessing techniques, and displaying them in a
structured and efficient manner.

3.2 Functional Requirements

These are the core operations the software must perform:
 Extract and manage image datasets.
 Load images from structured directories.
 Resize and preprocess images uniformly.
 Display images for visual verification.

3.3 Non-Functional Requirements

These define how the software should perform:
 The system should be efficient in memory usage to handle multiple images.
 Execution should be reliable and robust, handling corrupted or unreadable files gracefully.
 The software should provide quick visual feedback through plotted images.
 The code should be modular and easy to update or expand.

3.4 Software Tools Required

 Programming Language: Python 3.x
 Libraries: OpenCV for image processing, NumPy for numerical operations, Matplotlib for
visualization
 Environment: Jupyter Notebook or any Python IDE
 OS Compatibility: Windows, Linux, or macOS.

3.5 User Requirements

 The user should have basic knowledge of Python and Jupyter Notebook.

9
 The system must have the required libraries installed.
 The dataset should be organized or provided in a compressed format like ZIP.

3.6 Constraints
 Image formats must be supported by OpenCV (e.g., JPG, PNG).
 File paths and folder structures should remain consistent.
 The performance may vary based on the hardware capabilities of the user's system.

10
 Chapter-4
Software Design
The software design defines the structure and organization of the modules used in the image processing
project. A modular and layered approach was followed to ensure clarity, flexibility, and maintainability
throughout development.

4.1 Design Approach

The project adopts a modular design where each functional component is developed
independently. This allows easy debugging, future enhancements, and separation of concerns across the
image processing pipeline.

4.2 Dataset Handling Module

This module is responsible for managing and extracting the dataset. It ensures that the image
data is available in the correct directory structure, and handles any compressed files, such as ZIP
archives.

4.3 Image Loading Module

This module accesses the dataset and reads images from different category folders. It ensures all
valid images are loaded into memory while handling unreadable or corrupted files gracefully.

4.4 Image Preprocessing Module

This module resizes all images to a fixed dimension (224x224 pixels) for consistency. It also
performs color format conversions or normalization when required, preparing the data for visualization
or further machine learning tasks.

4.5 Visualization Module

The visualization module helps in plotting and inspecting images. It enables verification of
image quality, format, and preprocessing results, making it easier to spot any issues early in the
pipeline.

4.6 Flow of Control

The control flow of the software follows a linear sequence—from dataset extraction and image
loading, to preprocessing and visualization. This step-by-step execution ensures reliability and makes
the project easy to understand.

11
 Chapter-5
Implementation
5.1 Introduction
The design and implementation of the distracted driver detection system aim to create a robust,
real-time solution for identifying unsafe driving behaviors using in-vehicle camera images. By
leveraging a Convolutional Neural Network (CNN) based on MobileNetV2, the system processes
driver posture images to classify distractions with high accuracy and low inference time. This chapter
details the proposed system workflow, model architecture, preprocessing techniques, dataset, and
implementation setup, ensuring compatibility with edge devices for practical deployment in vehicles.

5.2 Proposed System

The proposed system follows a streamlined workflow to detect distracted driving behaviors:
 Image Capture: In-vehicle cameras capture real-time images of the driver.
 Preprocessing: Images are resized, normalized, and prepared for model input.
 Feature Extraction: The MobileNetV2 base model extracts relevant features from
preprocessed images.
 Classification: A custom classification head predicts the distraction class (e.g., texting, safe
driving).
 Alert Generation: If a distraction is detected, the system triggers an alert to warn the driver.

This workflow ensures efficient processing, enabling real-time performance with an inference time of
approximately 1 second per image, suitable for edge device deployment.

5.3 Model Architecture

The system employs MobileNetV2, a lightweight CNN optimized for mobile and edge
applications, with pre-trained weights from ImageNet to leverage transfer learning. The architecture is
designed as follows:
 Input Layer: Accepts RGB images resized to 160x160 pixels with 3 color channels.
 Base Model: MobileNetV2 with pre-trained ImageNet weights, where the base layers are
frozen to retain learned features.
 Custom Head:
o GlobalAveragePooling2D to reduce spatial dimensions.
o Dense layer with 512 units and ReLU activation for feature processing.
o Dropout layer (0.5) to prevent overfitting.
o Dense layer with 11 units and softmax activation for classifying 11 distraction classes.
12
The model is trained using sparse categorical crossentropy loss, the Adam optimizer (learning rate
0.001), and a batch size of 32. Training is conducted for approximately 20 epochs, with early stopping
applied if validation loss does not improve for 3 consecutive epochs.

This architecture achieves a validation accuracy of 96%, as detailed in Chapter 5, and is optimized for
real-time inference, with potential for further compression using TensorFlow Lite for edge deployment.

5.4 Preprocessing
Preprocessing is critical to ensure input images are compatible with MobileNetV2. The
following steps are applied:

 Resizing: Images are resized to 160x160 pixels to match the model’s input requirements.
 Normalization: Pixel values are normalized using MobileNetV2’s preprocess_input function,
which scales inputs to the range [-1, 1].
 Batching and Splitting: Images are processed in batches of 32, with a validation split of 0.2
(80% training, 20% validation).

No additional data augmentation (e.g., rotation, flipping) is applied, as the dataset’s diversity and
transfer learning suffice for robust performance.

5.5 Dataset and Implementation Setup

5.5.1 Dataset
The system is trained and evaluated using the State Farm Distracted Driver Detection
dataset, which contains 22,432 RGB images across 11 classes:

 c0: Safe driving

 c1: Texting (right hand)
 c2: Phone use (right ear)
 c3: Texting (left hand)
 c4: Phone use (left ear)
 c5: Operating radio
 c6: Drinking
 c7: Reaching behind
 c8: Hair and makeup
 c9: Talking to passenger
 c10: Other distractions

13
The dataset is split into 17,950 training images and 4,482 validation images, with each class
representing a distinct driver behavior. This diversity enables the model to generalize across
various distraction scenarios.

5.5.2 Implementation Setup

The implementation is developed using Python 3.12, TensorFlow 2.x, and supporting
libraries (NumPy, OpenCV, Matplotlib, Scikit-learn). Training and evaluation are performed on
a system equipped with an NVIDIA GTX 1660 GPU, ensuring efficient processing of the
dataset. The software environment is configured to handle large-scale image data and deep
learning workflows, with the model’s real-time inference capability tested to confirm a ~1-
second processing time per image. For edge deployment, the model is designed to be
compatible with optimization frameworks like TensorFlow Lite, although full deployment is
reserved for future work.

14
 References
[1] Abouelnaga, Y., Eraqi, H. M., & Moustafa, M. N., 2017. “Real-time distracted driver posture
classification,” arXiv preprint arXiv:1706.09498.

[2] Chen, X., Wu, C., Liu, Z., Zhang, N., & Ji, Y., 2021. “Computation offloading in beyond 5G
networks: A distributed learning framework and applications,” IEEE Wireless Communications, Vol.
28, No. 2, pp. 56–62.

[3] Cher, C., Dan, M., Karisa, H., Madonna, W., & Raby, M., 2015. “Using naturalistic driving data to
assess the prevalence of environmental factors and driver behaviors in teen driver crashes,” Technical
Report.

[4] Liu, D., 2020. “Driver status monitoring and early warning system based on multi-sensor fusion,”
Proceedings of the International Conference on Intelligent Transportation, Big Data, and Smart City
(ICITBS), pp. 24–27.

[5] Liu, Y., Zhang, Y., Li, J., Sun, J., Fu, F., & Gui, J., 2013. “Towards early status warning for driver
fatigue based on cognitive behavior models,” Proceedings of the International Conference on Digital
Human Modeling Applications in Health, Safety, Ergonomics, and Risk Management, Springer, pp.
55–60.

[6] National Highway Traffic Safety Administration (NHTSA), 2016. “2015 Motor Vehicle Crashes:
Overview,” Traffic Safety Facts, Research Note, Washington, DC, USA, pp. 1–9.

[7] Rasheed, F., Yau, K.-L.-A., Noor, R. M., Wu, C., & Low, Y.-C., 2020. “Deep reinforcement
learning for traffic signal control: A review,” IEEE Access, Vol. 8, pp. 208016–208044.

[8] World Health Organization (WHO), 2016. “Global Health Estimates 2015: Deaths by Cause, Age,
Sex, by Country and by Region (2000–2015),” Geneva, Switzerland.

[9] World Health Organization (WHO), 2017. “World Health Statistics 2017: Monitoring Health for
the Sustainable Development Goals (SDGs),” Geneva, Switzerland.

[10] Yau, K.-L.-A., Lee, H. J., Chong, Y.-W., Ling, M. H., Syed, A. R., Wu, C., & Goh, H. G., 2021.
“Augmented intelligence: Surveys of literature and expert opinion to understand relations between
human intelligence and artificial intelligence,” IEEE Access, Vol. 9, pp. 136744–136761.

15
[11] Rao, X., Lin, F., Chen, Z., & Zhao, J., 2019. “Distracted driving recognition method based on
deep convolutional neural network,” Journal of Ambient Intelligence and Humanized Computing,
https://doi.org/10.1007/s12652-019-01597-4.

16