📄 中文摘要
人类轨迹预测（HTF）旨在根据过去的轨迹和环境上下文预测未来的人类运动，广泛应用于自动驾驶、智能监控和人机交互等领域。尽管以往的研究关注于准确性、社会互动建模和多样性，但对不确定性建模、校准以及短观察期的预测关注较少，而这些因素对路径规划和碰撞避免等下游任务至关重要。DD-MDN是一种端到端的概率性HTF模型，结合了高位置准确性、校准的不确定性以及对短观察的鲁棒性。该方法使用少量去噪扩散骨干和双混合密度网络，学习有效的人类轨迹预测。通过这种方式，模型能够在短时间内提供可靠的轨迹预测，增强了在复杂环境中的应用潜力。
📄 English Summary
DD-MDN: Human Trajectory Forecasting with Diffusion-Based Dual Mixture Density Networks and Uncertainty Self-Calibration
Human Trajectory Forecasting (HTF) aims to predict future human movements based on past trajectories and environmental context, with applications in Autonomous Driving, Smart Surveillance, and Human-Robot Interaction. While previous research has focused on accuracy, social interaction modeling, and diversity, there has been limited attention to uncertainty modeling, calibration, and forecasts from short observation periods, which are critical for downstream tasks such as path planning and collision avoidance. DD-MDN is proposed as an end-to-end probabilistic HTF model that integrates high positional accuracy, calibrated uncertainty, and robustness to short observations. Utilizing a few-shot denoising diffusion backbone and a dual mixture density network, the method effectively learns human trajectory predictions. This approach enables reliable trajectory forecasting within short time frames, enhancing its applicability in complex environments.