A continuous spatio-temporal map of dynamics using implicit neural representations, achieving more accurate and efficient motion pattern representation and faster training.

Paper Code

An Map of Dynamics (MoD)-informed long-term human motion prediction framework (prediction horizon is 60 seconds) that incorporates a ranking method to output the most likely predicted trajectory, improving its practical utility for robotics applications. Evaluating on two real-world datasets showing that MoD-informed method outerperforms transformer-based and diffusion-based SOTA methods.

Paper Code

Online updating of human motion patterns to quickly adapt to changes, maintaining a probabilistic representation at each observed location and continuously updating parameters by tracking sufficient statistics with a stochastic Expectation-Maximization algorithm.

Paper Code

Using class attributes of heterogeneous agents to improve trajectory prediction accuracy; conducted a comparative study between motion pattern-based and deep learning-based trajectory prediction approaches on small, class-imbalanced datasets.

Paper Code

Present the Laminar Component Enhanced (LaCE) LHMP approach inspired by data-driven airflow modelling. LaCE-LHMP extracts laminar patterns in human dynamics and uses them for motion prediction, mitigating the impact of anomalous data in an unsupervised manner.

Paper Code

Present CLiFF-LHMP, which uses CLiFF-map, a specific MoD trained with human motion data recorded in the same environment to generate multi-modal trajectory predictions over extended periods of time.

Paper Code