MRA-FD: Driver Fatigue Detection via Deep Convolutional Modeling With Local and Global Features

Drowsy driving is a leading cause of traffic accidents, yet its detection remains challenging due to variations in head pose and the insufficient use of temporal information in traditional global face-based methods. Additionally, the scarcity of training samples affects the ability to learn effective fatigue features from images or videos. To address these challenges, the MRA-FD model is proposed, which integrates Multi-granularity feature extraction, feature Recalibration, and Attention-based feature fusion within a deep convolutional framework for Fatigue Detection. The model first employs a Multi-task Cascaded Convolutional Networks (MTCNN) to extract global facial features as well as multi-granularity local region features. The Feature Attention Module (FAM) then redistributes the weights of local stream components, while the proposed Global-Local fusion module effectively fuses global and recalibrated local features. To capture the temporal dynamics of fatigue, a Deep Bidirectional Long Short-Term Memory network (DB-LSTM) is employed to analyze long-term frame sequences. Furthermore, the IsoSSLMoCo pre-trained model alleviates the challenge of limited fatigue data annotations. Tested on a Raspberry Pi 4B, an edge computing device, the MRA-FD model demonstrates superior accuracy, achieving 90.10% and 94.37% on the Nthu-DDD and YawDD datasets respectively, which can cater well to the demand for real-time fatigue driving detection