All Projects Deep Learning · Image Classification · Road Safety

Distracted Driver Classification

Utilizing ResNet9 deep learning architecture to classify distracted driver behaviors, enhancing road safety — conducted at IIT Hyderabad under Prof. P. Rajalakshmi.

Dec 2023 – May 2024 IIT Hyderabad 99.22% Validation Accuracy

99.22%

Validation Accuracy

99.32%

F1 Score

Project Snapshot

Project Type

Deep Learning · Image Classification

Core Architecture

ResNet9

Dataset

State Farm Distracted Driver (Kaggle)

Key Metrics

Accuracy: 99.22% · F1: 99.32%

My Role

Research Intern

Supervisor

Prof. P. Rajalakshmi

Project Overview

The Challenge

Distracted driving is a significant contributor to road accidents globally. Effectively identifying and classifying these behaviors in real-time is crucial for enhancing road safety and developing preventative systems. This project addresses the challenge of accurately detecting various forms of driver distraction using computer vision techniques.

Our Solution

This project employs the ResNet9 deep learning architecture to classify ten distinct driver distraction classes from images. By leveraging the comprehensive State Farm Distracted Driver Detection dataset from Kaggle, and implementing robust image augmentation techniques, we developed a highly accurate and generalized model capable of contributing to safer driving environments.

Approach & Methodology

Deep Learning with ResNet9

Utilized the lightweight yet powerful ResNet9 architecture for its efficiency and ability to handle complex image features through skip connections, mitigating vanishing gradient issues.

Dataset & Preprocessing

Trained on the State Farm Distracted Driver Detection dataset comprising images across ten different driver behaviors (texting, talking on phone, safe driving, etc.).

Image Augmentation

Employed random rotation, cropping, and horizontal flipping to significantly enhance model robustness and generalization by exposing it to a wider variety of data.

Comparative Analysis (ResNet9 vs. ResNet50)

Conducted a comparative study demonstrating that ResNet9 achieves comparable high accuracy with a simpler architecture, making it suitable for deployment in resource-constrained environments.

Key Results & Visualizations

Accuracy vs Number of Epochs for ResNet9
Figure 1: Accuracy vs. Number of Epochs for ResNet9, showcasing rapid convergence and high performance.
Sample images tested on the ResNet9 model showing correct classifications
Figure 2: Sample Test Images with Predicted Classes, demonstrating the model's ability to accurately classify distracted behaviors.

Technologies Used

Python PyTorch ResNet9 Image Augmentation Convolutional Neural Networks (CNNs) Jupyter Notebook

Lessons Learned & Challenges

Data Imbalance

Addressing potential class imbalance within the dataset was crucial for achieving high and balanced performance across all distraction categories.

Overfitting Mitigation

Extensive use of image augmentation techniques proved vital in preventing overfitting, especially with a relatively smaller domain-specific dataset.

Architectural Choice

Demonstrating that ResNet9 achieves comparable results to much larger models (ResNet50) highlights the importance of selecting the right architecture — balancing accuracy with computational efficiency.

Impact of Skip Connections

Gained deeper understanding of how skip connections in ResNet effectively combat the vanishing gradient problem, enabling training of deeper networks.

Future Goals

Future work includes integrating the ResNet9 model into real-time distracted driver detection systems for smart vehicles and exploring additional deep learning architectures to further optimize performance and computational efficiency. Deployment on edge devices for in-car systems would be a valuable next step.

Read the Work

Published Research

This work has been published on ResearchGate. Read the full thesis.

Read Full Thesis All Projects