DTIU: A self-supervised grid-enhanced diffusion model for trajectory imputation in unconstrained scenarios

Abstract

While the widespread of location-based services has led to a proliferation of trajectory data, it is often accompanied by the inevitable problem of incomplete data due to various reasons, e.g., sensor failure and privacy concerns. Imputing the incomplete trajectory is critically important to a range of practical applications, e.g., traffic management and emergency response. Most recent proposals on trajectory imputation are designed for the urban scenario where a road network is available, which may not work well in the unconstrained environment scenario. As there are no road networks or predefined paths in the unconstrained environment, existing approaches for trajectory imputation may result in insufficient input data and thus lead to sub-optimal performance. To address this issue, we propose a self-supervised grid-enhanced Diffusion model for Trajectory Imputation in Unconstrained scenarios (DTIU). DTIU includes a diffusion model specifically designed for trajectory imputation tasks. DTIU avoids the problem of insufficient input by using trajectory data as the only input. To contend with missing position information and effectively learn the spatio-temporal correlations, we design a GridFormer based on AutoEncoder with an adaptive grid information enhancement strategy. DTIU adopts a self-supervised training strategy inspired by masked language models, aiming at address the data sparsity issue. Extensive experiments on real data offer insight into the effectiveness of the proposed framework. This marks the first application of diffusion models to tackle trajectory imputation in unconstrained scenarios.