The acquisition of trajectory generation methods based on GPS data is challenging due to privacy protection and high costs. This study proposes a method for generating vehicle trajectories using checkpoint data, which confronts several challenges: Firstly, the low checkpoint coverage results in discontinuous trajectories that are not compatible with existing models, and no research has been conducted filling the gaps. Secondly, existing models overlook road network constraints, preventing trajectories suitable for simulation. Lastly, these models lack the capability to produce diverse trajectories, which diminishes their practicality. To tackle these challenges, the TrajGAT-A* model is developed, which utilizes a graph neural network to build a costmap rich with actual traffic information, employs a clustering algorithm to create a functional area network and leverages a graph attention network to extract road network features. After constructing the costmap, the A* algorithm is applied to reconstruct continuous trajectories. Following this, the β-TrajVAE model is designed to segment the road network into intra and inter cluster sections via clustering algorithms and to conduct partitioned sampling. Hyperparameters are incorporated into the loss function to balance accuracy and divergence, leading to the generation of multiple costmaps for executing A* search and producing diverse trajectories. Experimental validation using Chongqing data reveals that the reconstructed trajectories surpass existing models in terms of Precision, Recall, and F1 metrics. The generated trajectories also outperform existing models in cross-entropy. Simulation experiments confirm that the proposed method is capable of generating trajectories that align with real traffic conditions.