Abstract: Capturing high-dimensional social interactions and trend features is essential for accurately predicting the feasible future behaviors of agents. To address this complexity, previous research has reduced the dimensionality of input variables through parametric curve fitting to capture more useful information, while other studies have inferred future trajectories using recursive or synchronous methods. However, these methods have limitations: a single smooth curve struggles to effectively fit social dynamics, recursive strategies can lead to cumulative errors, and synchronous strategies overlook constraints between future steps, rendering kinematic predictions infeasible. To overcome these challenges, a method combining singular value decomposition and feature pyramid networks is proposed to reduce dimensionality and extract trend features, eliminating redundant information. The proposed method replaces the traditional Euclidean space with a feature space based on singular value decomposition to better model multimodal predictions across different models. Results from various depths of trend feature predictions are progressively fused from the bottom layer to the top layer, generating final predictions through a global-to-local recursive trajectory prediction method. This recursive method employs interpolation techniques of varying granularity, integrating global information with the boundary information from each iteration??s region, continuously generating intermediate positional information for each area. Extensive experiments demonstrate that the proposed universal trajectory prediction framework significantly enhances the prediction accuracy and reliability of existing trajectory models on public benchmarks.

Key words: singular value decomposition, feature pyramid network, recursive trajectory generator, feature extraction, multimodal, trajectory prediction

摘要： 捕捉高维社会活动和趋势特征对准确预测智能体的可行未来行为至关重要。为应对这一复杂性，已有研究通过参数曲线拟合降低输入变量的维度，以捕获更有用的信息；而另一些研究则采用递归或同步方式推断未来轨迹。然而，这些方法存在一些不足之处：单一的平滑曲线难以有效拟合社会动态，递归策略易导致累计误差，而同步策略则忽略了未来步骤之间的约束，进而使运动学上的预测变得不可行。为了解决这些问题，提出了一种结合奇异值分解和时间序列特征金字塔网络的方法，旨在降维和提取趋势特征，以去除冗余信息。该方法采用基于奇异值分解的特征空间替代传统的欧几里德空间，以在该空间内模拟不同模型的多模态预测。从底层到最上层逐步融合不同深度趋势特征的预测结果，并通过全局到局部的递归轨迹预测生成方法生成最终预测结果。该递归轨迹生成方法使用不同粒度的插值技术，将全局信息与每次迭代区域的头尾部信息相结合，持续生成每个区域的中间步骤位置信息。大量实验证明，所提出的通用轨迹预测框架显著提高了现有轨迹模型在公共基准上的预测精度和可靠性。

关键词: 奇异值分解, 特征金字塔网络, 递归轨迹生成器, 特征提取, 多模态, 轨迹预测