Detection of Fine-grained Driver Cognitive Responses for Cross-subject Deep Learning Framework

Understanding driver's neural response to unexpected events is critical to developing a reliable brain-based driver monitoring system. This study investigates the neuro-physiological properties of braking behavior in sudden driving situations using brain waves recorded in real road environments. Two types of unexpected events have been designed: sudden deceleration of a front vehicle and sudden appearance of a pedestrian dummy. Event-related potentials (ERPs), including N200 and P300, and Electroencephalography (EEG) microstate dynamics were analyzed to characterize drivers' cognitive and behavioral responses. To classify brake and rest-state, we propose a multi-branch classification framework that utilizes microstate-based time dynamics and ERP-based summary features together. Microstate label sequences and global field power learned state transition patterns through recurrent neural networks, and ERP features extracted phase-related neural responses through the mediated sub-network. The two representations were fused and utilized in the classification process, and their performance was evaluated by the leave-one-subject-out (LOSO) protocol. The proposed model achieved an average accuracy of 0.9364 and a macro F1-score of 0.5755 under front vehicle deceleration conditions. Similar performance was shown in pedestrian appearance conditions, and stable generalization was confirmed throughout the subjects and driving conditions. These results demonstrate that the way microstate dynamics and ERP features are combined can more reliably capture braking-related neural responses, highlighting the usefulness of approaches utilizing complex neural signals in EEG-based driver intention recognition.