Attention-based deep neural network for driver behavior recognition

Abstract

Driver behavior recognition is crucial for traffic safety in intelligent transportation systems. To understand the driver distraction behavior, deep learning methods has been used to learn the hierarchical features of receptive field images. However, most of existing studies focus on investing in spatial components based on the convolutional neural network for visual analysis. In this study, an attention-based deep neural network (ADNet) method is proposed for driver behavior recognition. The ADNet framework is first presented, which integrates the lightweight attention block into the deep learning model to strengthen the representative power of the model. Then, the channel attention (CA) block is designed to model inter-channel dependencies in the ADNet. Furthermore, the spatial attention block is combined with the CA for adaptive feature extraction. Data augmentation is used in the data preprocessing stage to improve recognition performance. The effectiveness of the proposed method is verified on the distraction dataset of American University in Cairo (AUC) and the real dataset of Hunan University (HNU). The ADNet method achieves a Top-1 accuracy of 98.48%, which outperforms the state-of-the-art methods.