A lightweight model for distracted driver detection based on neural architecture search and coordinate attention

Research indicates that driver distraction is a leading cause of most accidents. The distracted driver recognition task can be modeled using deep learning, and the lightweight deep learning model has been extensively studied for this task. In this study, we designed a novel lightweight network by combining a neural architecture search and the coordinate attention mechanism (CA) based on the characteristics of the distracted driver recognition task. The designed neural architecture searching space was searched to find the most suitable network for recognizing distracted driver behaviors. This model, combined with CA, forms a new lightweight framework specifically designed for distracted driver recognition, called CondConv coordinate attention network (CCNET). Experimental evaluations were conducted on two public datasets: AUC distracted driver and State Farm distracted driver detection. The accuracy of CCNET for AUC was 95.52%, whereas it ranged from 99.8% to 99.9% for the State Farm. To simulate real driving scenarios, the model was deployed on an embedded device (Nvidia Jetson Nano B01), achieving a real processing speed of 54 frames per second and enabling real-time processing. These experimental results indicate that the performance of CCNET is superior to that of most previous lightweight models. Notably, the CCNET model has only 1.7 million parameters and an efficient inference speed, making it smaller and faster than most previous lightweight models.