AIMSafe: EEG-Based Driver Behavior Understanding via Attention and Incremental Learning Mechanisms

In this paper, we propose AIMSafe, an electroencephalographic (EEG) based system that studies driver in-vehicle behaviors leveraging Attention networks and Incremental learning Mechanism for road Safety. Instead of using predefined classes, we categorize driver in-vehicle activities into different risk levels - a stronger motion may have a higher chance of unsafe driving. More specifically, we first employ a CNN based model to distinguish two basic activities - 1. normal driving and 2. unsafe driving. Moreover, AIMSafe also leverages smartphone IMU sensors generating soft hints that helps automatically label EEG data on road. We then adopt class-incremental learning to rank other Out-of-Distribution (OOD) driver activities (safe to unsafe) based on the Mahalanobis distance. A modified Squeeze-and-Excitation (SE) block is also used to adaptively select effective EEG electrodes for improving the system efficiency. Evaluation results (involving 11 males and 4 females) show that AIMSafe could achieve a detection accuracy over 95% on unsafe driving activities using only 4 electrodes.