Temporal Convolutional Networks Reconstruct Missing Trajectory Data

A new research paper introduces a method for "trajectory inpainting," which involves reconstructing missing portions of movement data. This is crucial because real-world trajectory data often suffers from gaps due to sensor malfunctions or communication issues. The proposed solution utilizes a Temporal Convolutional Network (TCN) that employs symmetric dilation, allowing it to consider both preceding and subsequent data points when filling in the gaps. This approach differs from typical forecasting models that only look backward. The TCN is trained with a specialized loss function that combines weighted mean squared error, penalties for boundary discontinuities, and a smoothness regularizer. This comprehensive training strategy helps the model generate coherent and realistic missing segments. Experiments on a synthetic dataset of two-dimensional trajectories, where 20% of segments were randomly masked, demonstrated that the model achieved good R-squared, Mean Squared Error (MSE), and Mean Absolute Error (MAE) metrics, indicating its effectiveness in accurately inferring the missing data.