MAViT: A lightweight hybrid model with mutual attention mechanism for driver behavior recognition

Driver behavior plays a critical role in transportation safety. Studies show that over half of traffic accidents are caused by distracted driving, making it essential to accurately recognize various driver behaviors. These behaviors can be reflected by their overall postures and local details, both of which need to be considered. A new lightweight hybrid model, Mutual Attention Vision Transformer (MAViT), is proposed in this paper with a mutual attention mechanism to recognize these behaviors. In this model, a multi-scale inverse residual (MSIR) block is designed to extract key features of different scales from the driving image. In addition, the local and global mutual attention module (LGMA) is proposed, providing stronger representation capabilities in driving scenarios by establishing the mutual influence relationship between local features and global features. The proposed model utilizes only 0.2 million parameters and achieves accuracy rates of 99.64% and 99.40% on the State Farm Distracted Driver Detection (SFD3) and the American University in Cairo Distracted Driver v2 (AUCv2) public datasets, respectively. This surpasses many other state-of-the-art methods.