The driving styles of the human vehicles (HVs) in the lane-changing target lane are incomplete information for the game modeling process of the autonomous vehicles (AVs), making it difficult for AVs to predict the driving decisions of HVs. A Bayesian game lane-changing prediction model is proposed, which solves incomplete information among game participants. First, vehicle acceleration’s subjective perception evaluation model is used to obtain the initial quantified value of HVs’ driving styles in the target lane, which is used for the state space’s confidence of HVs in the Bayesian game lane-changing model. Secondly, the expected safety distance based on the molecular dynamics model is used to solve the game safety benefits. Finally, the Bayesian game is transformed into a complete information static game to solve the Nash equilibrium. By comparing the matching degree of HVs’ predicted actions with actual actions, combined with the quantitative indicators of driving styles, the state space’s evaluation confidence is modified. The simulation results show that compared with the K-level game model and the robust K-level game improvement model, the Bayesian game model has a collision rate of 2% in the same traffic environment by using more stringent collision rate statistics standards, which is far lower than the 13% of the K-level game and the 7% of the robust model. The three game models account for 93%, 92% and 69% of lane-changing within 3.5 s, respectively. The proposed lane-changing model has a better effect on the decision-making risk control of AVs, and the efficiency of lane-changing within a reasonable time can also be guaranteed.