TrajDiff: A method for vessel trajectory imputation utilizing resampled conditional diffusion models

Automatic Identification System (AIS) data plays a key role in maritime monitoring but often experiences missing trajectory points due to signal loss, equipment failures, or intentional non reporting. These gaps compromise the accuracy of maritime analysis. Existing imputation methods struggle with long gaps and fail to capture global spatiotemporal patterns. Machine learning approaches, while promising, introduce high computational complexity. This paper presents a cyclic denoising trajectory imputation method using conditional diffusion models. The method generates latent AIS data representations by adding noise and performing reverse reconstruction. This process enhances learning of trajectory distributions and motion features. Key contributions include: (1) a vessel trajectory diffusion model that uses noise-based denoising, (2) a conditional embedding mechanism that uses known trajectory points to guide generation, and (3) a Cyclic Denoising Real Trajectory Embedding (CDN-RTE)module that integrates known and generated trajectories to restore details. The model was tested on public datasets, outperforming baseline methods in various missing data scenarios. It reduced 10-hour trajectory imputation Root Mean Square Error (RMSE) by at least 13.9%. The results demonstrate its effectiveness in improving maritime situational awareness and security. • Diffusion mode for long-term vessel trajectory imputation. • Embedding encodes spatiotemporal data, guides trajectory generation. • Resampling refines details, enhances recovery accuracy.