Long-Term Human Motion Prediction Using Spatio-Temporal Maps of Dynamics

Yufei Zhu, Andrey Rudenko, Tomasz Piotr Kucner, Achim J. Lilienthal, Martin Magnusson

Introduction

Predicting human movement over short horizons—typically 2 to 5 seconds—is a well-studied problem in robotics, often solvable by extrapolating current velocity and heading. However, Long-Term Human Motion Prediction (LHMP), which looks up to 60 seconds into the future, presents a significantly more complex challenge. At these extended intervals, individual intent becomes secondary to the physical constraints of the environment and established social conventions.

Standard deep learning approaches, such as LSTMs and Transformers, frequently struggle with LHMP. Because they rely heavily on recent trajectory sequences, they often lack a global understanding of the environment, leading to "unfeasible" predictions where agents appear to walk through walls or obstacles. This paper introduces MoD-LHMP, a framework designed to bridge this gap by integrating "Maps of Dynamics" that account for both spatial layouts and temporal shifts in human traffic.

The MoD-LHMP Framework

The core philosophy of MoD-LHMP (Map of Dynamics-informed Long-Term Human Motion Prediction) is that the environment dictates long-term behavior more than immediate momentum. Instead of treating prediction as a pure sequence-to-sequence problem, the framework queries a map that stores historical movement patterns for specific locations.

Time-Conditioned CLiFF-maps (TC-CLiFF)

To capture the nuances of human movement, the authors enhanced the Circular-Linear Flow Field (CLiFF) map. Traditional maps often aggregate data across an entire day, which obscures time-dependent patterns like morning rushes or lunch breaks. The proposed TC-CLiFF creates distinct maps for different time intervals. This allows the model to recognize that a hallway might primarily flow in one direction at 9:00 AM and the opposite direction at 5:00 PM.

Trajectory Ranking and Filtering

Because human motion is inherently stochastic, the model generates multiple potential future paths. To identify the most plausible outcome, MoD-LHMP employs a Trajectory Ranking system. Each path is assigned a "fitness score" based on its alignment with the learned dynamics of the TC-CLiFF map.

Furthermore, the framework utilizes Kernel-Based Filtering to balance a person's current momentum with the map's suggestions. By applying a Gaussian kernel, the model can dynamically weight the map's influence. If an individual's current movement strongly contradicts the historical map (an outlier), the model reduces the map's weight to accommodate the unique behavior.

Experimental Evaluation

The framework was rigorously tested on two large-scale, real-world datasets:

  1. ATC Dataset: Collected over 92 days in a Japanese shopping mall, featuring high-density traffic and complex spatial topologies.
  2. Edinburgh Dataset: Collected over several months in a university building, capturing periodic flows between classes and "idling" behaviors where people congregate in specific areas.

Key Results and Insights

The MoD-LHMP framework demonstrated significant performance gains over state-of-the-art deep learning baselines, such as Trajectron++:

  • Accuracy: In the ATC dataset, the model achieved a 50% reduction in Average Displacement Error (ADE) for 60-second horizons. This improvement stems from the model's inherent respect for environmental boundaries—unlike pure DL models, MoD-LHMP rarely predicts trajectories that intersect with physical obstacles.
  • Temporal Awareness: The inclusion of time-conditioning in TC-CLiFF proved vital. Accounting for the time of day significantly lowered error rates compared to static maps, confirming that human motion is a spatio-temporal phenomenon.
  • Computational Efficiency: Despite its accuracy, the system remains highly efficient. It achieves an average inference time of 0.11 seconds for a 60-second prediction on a standard laptop CPU, making it suitable for real-time robotic deployment.

Limitations and Future Directions

While MoD-LHMP represents a major step forward, the authors note several areas for refinement. The current use of a uniform 1m x 1m grid can result in "blocky" transitions; moving toward continuous coordinate mapping could improve smoothness. Additionally, the framework currently uses a fixed confidence parameter to balance map data against current velocity. Future iterations may involve learning this parameter online to increase adaptivity. Finally, while the model excels at environment-aware navigation, integrating explicit person-to-person social avoidance (social forces) would further enhance its realism in crowded spaces.

Conclusion

MoD-LHMP demonstrates that for long-term horizons, the "topology" of a space is the most reliable predictor of human motion. By combining time-conditioned flow fields with a robust ranking mechanism, the framework provides a computationally efficient solution that keeps robotic predictions "on the rails" and contextually grounded.

Summary updated

BibTeX
@article{zhu2025long,
  author = {Yufei Zhu and Andrey Rudenko and Tomasz Piotr Kucner and Achim J. Lilienthal and Martin Magnusson},
  title = {{Long-Term Human Motion Prediction Using Spatio-Temporal Maps of Dynamics}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2025},
  doi = {10.1109/lra.2025.3619831},
}
Figure