Dual-view diffusion for pedestrian trajectory imputation

Trajectory imputation aims to forecast the missing values in the observed trajectories. Factors such as sensor failures and occlusions can frequently result in missing coordinates in observed trajectories. For real-world applicability, trajectory imputation models must effectively impute these missing observations, thereby improving the performance of trajectory prediction models. Only a few works focus on addressing the trajectory imputation task. To this end, we propose a novel Synchronous Dual-View Diffusion, a score-based diffusion model, to effectively address missing values in trajectories. The core of the proposed framework is to leverage the diffusion model conditioned on a dual view of the observed trajectory, where both views are synchronously utilized for the imputation task. Additionally, we impose similarity constraints in the imputation process by maximizing the alignment between predictions from the dual views. This enables better imputation of trajectory samples consistent with the ground truth trajectory. Extensive experimentation on benchmark datasets demonstrates that our approach consistently improves imputation performance compared to existing imputation methods. Furthermore, our approach demonstrates consistent improvements over strong baselines across five pedestrian trajectory benchmarks under both easy and hard missingness protocols, achieving best performance with up to 81.1% lower MAE on ZARA1-M and 72.3% lower RMSE on ZARA2-M compared to the second-best method. Code is publicly available at https://github.com/Pranav-chib/SyncDVD.