Autonomous driving systems have made great strides in danger perception, but they often ignore the driver’s intuitive perception of danger, unintentionally reducing driving comfort and trust in these systems. This oversight highlights the necessity of aligning technological responses with human perception to enhance the overall safety and efficacy of automated driving. Considering the complexity and variability of real-world driving, our research introduces the Spatio-Temporal Graph Convolution Network for Hazard Perception, namely STGCN-DHD. This method combines dynamic brain network, graph convolution, and multi-head self-attention mechanism, improving the spatio-temporal resolution of neural data analysis, enabling STGCN-DHD to effectively capture the rapid and complex neural interactions in driving hazard perception. Validated on two datasets related to driving hazard perception, our approach achieves top recognition rates of 86.69% and 72.36%, surpassing other models by 7.29% and 4.98% respectively, and demonstrates robust performance even with reduced data input. Our methodology underscores the importance of considering drivers’ intuitive responses in automated driving system design and offers a novel approach for enhancing driving hazard perception. It expands the applicability and practicality of Electroencephalography (EEG) and neural network models in real-world settings, potentially poised to enhance the synergy between human intuition and machine automation.