LiteDriveNet for driver distraction classification using a lightweight multi-scale convolutional neural network

Distracted driving is a critical global concern and a major factor in numerous road collisions. To mitigate these issues, this article introduces a novel model, LiteDriveNet, a resource-efficient 16-layer convolutional neural network with a compact size of 1.71 MB, designed for precise image identification and behavioral analysis. The architecture incorporates multi-scale receptive blending, progressive feature enhancement, and hierarchical feature aggregation to achieve robust feature extraction. To assess the model performance, a dataset named DistractedDrivingSet_v1 is also proposed, comprising 6075 outdoor images across 8 distinctive classes, captured under real-world lighting conditions such as sunlight and shadows. LiteDriveNet's efficacy was further verified using two publicly available benchmark datasets, the State Farm Distracted Driver Detection (SFD3) dataset and the American University in Cairo (AUC) version 2. Experimental findings show that LiteDriveNet consistently surpasses both accuracy and computational efficiency compared to existing state-of-the-art, lightweight, and neural architecture search-based (NAS) models. LiteDriveNet achieves average validation accuracy improvements of 16.53% on DistractedDrivingSet_v1, 6.70% on SFD3, 10.91% on AUCv2 Camera1, and 28.97% on AUCv2 Camera2. An ablation study also proves that LiteDriveNet is an eminent lightweight model for distracted driver recognition, as it is highly suitable for real-time environments compared to state-of-the-art techniques.