A Lightweight and Efficient Distracted Driver Detection Model Fusing Convolutional Neural Network and Vision Transformer

Identifying distracted drivers is crucial for enhancing driving safety and advancing intelligent driver assistance systems. Recently, researchers have applied Convolutional Neural Network (CNN) and Vision Transformer (ViT) models for driver state decision. However, both models often suffer from several issues such as numerous parameters and low detection efficiency. To address these challenges, this study proposes the Convolution Vision Transformer (CoViT) model for distracted driver identification, leveraging techniques such as Low Complexity Attention Mechanism (LCAM), Multi-scale Dilation Convolution (MSDC), and Depth Separable Convolution (DSC). Moreover, the CoViT model features a typical "pyramid" structure, enabling effective feature extraction across different scales. Subsequently, the proposed system is trained and evaluated using the publicly available driving behavior datasets SFD2 and 100-Driver, as well as real-world road experiments. Experimental results show that the CoViT model yields high recognition performance, with mean Accuracy (mAcc) scores of 95.17%, 97.89%, and 93.54% on the recorded dataset, SFD2 dataset, and 100-Driver dataset, respectively. These scores surpass those obtained by similar lightweight models. Furthermore, ablation experiments reveal that deep and dilated convolution significantly enhance model performance. In addition, the CoViT model demonstrates its applicability to real-time driving behavior detection tasks, with a parametric count of just 1.24M - a reduction of 2.67M compared to MobileNetV3 - and an online inference Frames Per Second (FPS) of 159.13.