In the information society, it is very important to discover and predict human position trajectories and activity patterns based on users' historical activity trajectories. Most of the existing studies use the Markov model based on the similarity classification between time and trajectory, ignoring the differences of movement laws under different travel modes. This paper distinguishes different travel modes, based on the characteristics of the trajectory's speed, acceleration, and heading change speed, the XGBoost algorithm is used to identify the travel mode of trajectory, and the optimized trajectory segmentation algorithm is used to decompose the human travel trajectory into travel modes. Markov models trained by trajectories of different travel modes are used to accurately predict travel trajectories. Experiments show that there are significant differences in the trajectory movement laws of different travel modes, and the prediction accuracy and distance deviation of the proposed method are obviously better than several benchmark methods.