Trajectory data is vital for AI-driven urban computing, yet its publication poses severe privacy risks. Existing differential privacy schemes fail to balance utility and protection. We propose PerTrajTree-DP, a novel framework unifying personalized sensitivity detection with hierarchical protection. First employs a TF-IDF mechanism to automatically identify sensitive locations, followed by a customized perturbation based on sensitivity levels. The perturbed trajectories are then encoded via Hilbert curves and aggregated into a privacy-aware prefix tree. A layered budget allocation strategy injects noise into node counts, preserving structural integrity while ensuring privacy. Experiments on a real-world dataset show PerTrajTree-DP consistently outperforms baseline methods, reducing query MAE by up to 40% under moderate privacy budgets. Our work provides a practical and effective solution for publishing high-utility, trustworthy trajectory data for AI applications.