Name: MARCELO BRINGUENTI PEDRO
Publication date: 16/12/2024
Examining board:
Name |
Role |
|---|---|
| ALBERTO FERREIRA DE SOUZA | Presidente |
| CLAUDINE SANTOS BADUE | Examinador Interno |
| FRANCISCO DE ASSIS BOLDT | Examinador Externo |
Summary: Driver distraction is a leading cause of traffic accidents, with the World Health Organization reporting over 40,000 annual fatalities in Brazil and studies indicating that up to 80% of accidents are linked to driver inattention. To address this critical issue, we propose DeepDDBD, an automated system for detecting and classifying distracted driving behaviors using deep neural networks. The proposed system introduces three key innovations: (1) a novel Adaptive Split-Fusion (ASF) architecture that integrates Convolutional Neural Networks (CNNs) with Transformer models to enhance feature extraction and classification accuracy, (2) validation across three datasets, including a custom dataset designed to capture a wide range of distraction scenarios, and (3) classification of ten distinct driver behaviors encompassing cognitive, visual, and manual distractions. Experimental results demonstrate that DeepDDBD achieves an average accuracy of 90,14% across all datasets, outperforming baseline models such as EfficientNetV2-M and ResNet152v2. The system excels in distinguishing between attentive driving and nine different distracted behaviors, maintaining robust performance across diverse environmental conditions and driver profiles. The primary contributions of this work include the development of the DeepDDBD architecture tailored for real-time driver behavior monitoring, the release of a custom dataset for future research, and empirical validation of the ASF approach. These advancements provide a foundation for practical applications in driver monitoring systems aimed at improving road safety.
