Based on the vehicle trajectory data of urban underground roads in Shanghai, this paper proposes a theory-data driven vehicle car-following model in urban underground roads. The vehicle trajectory data are collected and processed by millimeter wave radar to screen the car-following behavior. Based on the characteristics of car-following behavior on underground roads, a theory-driven intelligent driver model (IDM) and a data-driven long short-term memory (LSTM) car-following model are developed for acceleration, velocity, and distance headway, and error comparison is conducted. Finally, the prediction results of the two types of models are regarded as two sets of observations for the same system state obtained by different observation methods, and the real driving state is predicted by the adaptive Kalman filter method. The IDM-LSTM hybrid model based on speed prediction is developed, achieving a root mean square error of 0.118 7, outperforms the single IDM (0.583 6) and LSTM (0.123 9) models. This paper can provide valuable reference for the operation and road safety management of urban underground roads.