Driver distraction stands out as a leading cause of traffic accidents. While existing scene graph risk assessment models only consider the spatial relationships and interactions between traffic participants, they often fall short in mirroring human-like scene comprehension—particularly in dynamically allocating and processing the complexity of human visual attention. Inspired by the human attention mechanism, we propose an innovative dynamic attention-enhanced spatio-temporal model for pedestrian collision risk assessment. In particular, our proposed model utilizes the traffic scene graph to capture the spatial relationships of ego-vehicle and pedestrians. Besides, we introduce a dynamic attention enhancement model that simulates the dynamic allocation of driver visual attention, achieving an advanced spatio-temporal understanding of traffic scenes. The proposed network has been validated on multiple datasets, including IESG, Non-IESG, and the 1043-Carla datasets. Comparative experiments against other state-of-the-art methods demonstrate the superior performance of our network in subjective risk assessment tasks.