<p>Local differential privacy (LDP) is one of the important technologies for protecting the security and privacy of trajectory data. However, the existing local differential privacy algorithms usually rely on a single spatial dependency method to evaluate the sensitivity of trajectory data. These methods fail to fully consider the differential impact of overall encoding perturbation and regional encoding perturbation on data utility, as well as the personalized requirements for trajectory data encoding perturbation. To address this issue, we introduce the three-way decisions (3WD) and optimize the LDP models by improving the quadtree-based location perturbation model (QLP) and the quadtree-based joint location perturbation model (QJLP), proposing two new models: the three-way decision-based QLP model (3WDQLP) and the three-way decision-based QJLP model (3WDQJLP). The two models differ in the encoding perturbation process. Specifically, we generate 3WD results using different spatial dependency methods and derive various positive and negative region results based on K-means-based adaptive clustering, analyzing the dependency degree of trajectory data from multiple perspectives. Then, we integrate these results using a label fusion strategy to generate new 3WD results. Subsequently, we divide the positive and negative region data into multiple equivalent sets using hierarchical clustering and further segment the boundary region into positive and negative regions using centroid clustering for encoding perturbation separately. Finally, we conduct experiments on eight datasets using the proposed models. The experimental results fully validate that the proposed models demonstrate significant advantages over existing local differential privacy models in terms of error suppression, classification accuracy, and resistance to attacks.</p>