Equipping with Human Cognition: Driver Intention Recognition with Multimodal Information Fusion.

The article focuses on developing a driver intention recognition system that integrates multimodal cognitive information—including electroencephalogram (EEG) signals, eye movement data, and driving operation data—to enhance autonomous driving systems' perception and decision-making in complex urban environments. Using a driver-in-the-loop simulation platform with hazardous scenarios, the study collected and preprocessed multimodal data, applying feature-level fusion to train deep learning models based on multilayer perceptron (MLP), convolutional neural network (CNN), and Transformer architectures. Results show that the Transformer-based model with multimodal fusion achieved the highest accuracy (93.02%) in predicting emergency braking and steering evasion intentions, outperforming models using single modalities or simpler architectures. The research highlights the complementary value of combining physiological and operational data for driver intention recognition, while discussing challenges such as computational demands, privacy concerns, integration with existing vehicle sensors, and the need for further validation in real-world conditions.