Abstract Driver distraction is a major factor contributing to traffic accidents, making its automatic detection crucial for enhancing road safety. Convolutional Neural Networks (CNNs) have demonstrated considerable potential in addressing this problem, but they face significant challenges, including high computational costs, feature redundancy, insufficient attention to critical areas, and limited generalization ability. In this paper, we propose a novel deep convolutional network designed to overcome these challenges by incorporating Autoencoder-based feature dimensionality reduction, an attention mechanism to improve feature selection, and Dropout layers to enhance generalization and prevent overfitting. The convolutional layers capture spatial features from driver images, and the attention mechanism prioritizes the most relevant areas that are indicative of distraction behaviors. Dropout layers are integrated during training to improve model robustness and prevent overfitting. A key contribution of our Autoencoder-based Hybrid Deep Convolutional Neural Network (AHDCNN) is the integration of an Autoencoder that reduces the dimensionality of the extracted features, minimizing redundancy and improving computational efficiency. Distraction behaviors are classified using Fully Connected Layers combined with a Softmax classifier. Experimental results using the StateFarm dataset demonstrate high accuracy with reduced computational complexity, making the proposed method particularly well-suited for real-time driver monitoring systems, ultimately contributing to improved road safety.
Hassam et al. (Mon,) studied this question.