Research Analyzer
← Back ICRA 2026

DD-MDN: Human Trajectory Forecasting with Diffusion-Based Dual Mixture Density Networks and Uncertainty Self-Calibration

Manuel Hetzel, Kerim Turacan, Hannes Reichert, Konrad Doll, Bernhard Sick

PDF

AI summary

Key figure (auto-extracted from paper)
DD-MDN achieves state-of-the-art trajectory accuracy while delivering self-calibrated uncertainty estimates directly from short observation windows.
Human Trajectory Forecasting Uncertainty Calibration Diffusion Models Mixture Density Networks Probabilistic Forecasting Autonomous Systems

Problem

Existing trajectory forecasting models prioritize positional accuracy but neglect reliable uncertainty quantification, leading to misplaced confidence that compromises downstream safety-critical tasks.

Approach

The model integrates a few-shot denoising diffusion backbone with a dual mixture density network to learn self-calibrated uncertainty and generate probability-ranked trajectory hypotheses without predefined anchors.

Key results

  • State-of-the-art accuracy across ETH/UCY, SDD, inD, and IMPTC datasets
  • Self-calibrated aleatoric uncertainty without post-hoc recalibration
  • Robust forecasting performance from short observation intervals
  • Probability-ranked trajectory hypotheses without predefined waypoints

Why it matters

Enables safer autonomous decision-making by providing trustworthy confidence estimates alongside accurate predictions for path planning and collision avoidance.

Abstract

Human Trajectory Forecasting (HTF) predicts future human movements from past trajectories and envi- ronmental context, with applications in Autonomous Driving, Smart Surveillance, and Human-Robot Interaction. While prior work has focused on accuracy, social interaction modeling, and diversity, little attention has been paid to uncertainty modeling, calibration, and forecasts from short observation periods, which are crucial for downstream tasks such as path planning and collision avoidance. We propose DD-MDN, an end-to-end probabilistic HTF model that combines high positional accuracy, calibrated uncertainty, and robustness to short observations. Using a few-shot denoising diffusion back- bone and a dual mixture density network, our method learns self-calibrated residence areas and probability-ranked anchor paths, from which diverse trajectory hypotheses are derived, without predefined anchors or endpoints. Experiments on the ETH/UCY, SDD, inD, and IMPTC datasets demonstrate state- of-the-art accuracy, robustness at short observation intervals, and reliable uncertainty modeling. The code is available at: https://github.com/kav-institute/ddmdn.

Index terms

Planning under Uncertainty Probabilistic Inference Deep Learning Methods

Related papers