In order to further analyze the visual characteristics and risk conditions of drivers at the confluence sections of entrance ramps at adjacent urban underground roadways, data collected from real-vehicle experiments were used. A linear fitting model was constructed and the pupil area growth rate was calculated to investigate the patterns of pupil area changes. The K-medoids clustering method was employed to classify the regions of interest in drivers' gaze patterns, and the characteristics of drivers' gaze behavior were analyzed. A game theory-extension cloud evaluation model was constructed to evaluate the driving risks at the confluence sections of entrance ramps at adjacent urban underground roadways. The results show that the pupil area increases linearly in the entrance section and decreases linearly in the exit section. The overall load follows the order of adjacent entrance confluence section > adjacent entrance split section > adjacent exit split section > adjacent exit confluence section. Additionally, drivers face certain driving risks due to various factors at the confluence sections of entrance ramps at adjacent urban underground roadways. Based on the evaluation and actual survey, the risk factors for each section were analyzed, and optimization and improvement suggestions were proposed.