{
  "paper_id": "b8bbfd322f0187520b23bd872d5a2e2c",
  "paper_title": "Human Trajectory Prediction Using Similarity-Based Multi-Model Fusion",
  "paper_authors": [
    "Golnaz Habibi",
    "Jonathan P. How"
  ],
  "publication_date": "",
  "publication_venue": "IEEE Robotics and Automation Letters",
  "abstract": "Understanding pedestrian behaviors is crucial for a safe navigation of self-driving vehicles. However, pedestrians exhibit a large variety in their motion behaviors that are affected by interaction with the environment and other users of the road/sidewalk. Most of current models are limited to batch (offline) settings which requires to learn from the entire datasets. This letter presents a similarity-based model fusion algorithm, called SimFuse, for improving the prediction accuracy which enables autonomous agents to incrementally update their knowledge by communicating with other vehicles (V2V) by infrastructures (V2I). In the proposed framework, knowledge is shared in a compact form of “learned model” instead of raw data which provides a scalable sharing paradigm between agents. This work extends our prior work SILA [^1] by providing multi model fusion of n ≥ 2 models at the same time. We evaluate our algorithm in both intersection and non intersection scenarios and compare it with other baselines. The results show our algorithm outperforms state of the art in terms of Average Displacement Error at intersection scenarios and it has comparable result for non intersection scenarios with 3% improvement over SILA in ADE. The results also show that SimFuse updating time is up to 12 times faster than SILA with similar performance.",
  "keywords": [
    "Distributed Robot Systems",
    "Human and Humanoid Motion Analysis and Synthesis",
    "Computer Vision for Transportation"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:37.285775Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文提出一种基于相似度的多模型融合算法SimFuse，用于提高行人轨迹预测的精度。该算法解决了现有模型通常依赖批量离线学习的局限，使自动驾驶车辆能够通过V2V/V2I通信以增量方式更新知识。相关工作如SILA只能融合两个模型，而SimFuse支持同时融合多个模型，通过将知识以‘学习模型’的紧凑形式共享，而非原始数据，实现可扩展的共享范式。方法包括三步：计算运动原语间的余弦相似度、构建匹配图并通过边松弛实现一致性（确保两个模型间一对一的匹配）、融合匹配的运动原语及其过渡（用稀疏高斯过程融合多个GP）。在数据集上，采用ETH、Hotel、Univ、Zara1、Zara2等非交叉场景以及自采集的交叉路口场景，使用leave-one-out方法，指标为平均位移误差（ADE）和最终位移误差（FDE）。实验表明，SimFuse在交叉路口场景中相比最先进方法ADE提升30%，在非交叉场景平均ADE比SILA提升3%，且融合速度比SILA快达12倍（融合35个模型时）。未来工作可扩展至其他交通参与者（如自行车和汽车），结合SILA的增量学习实现更灵活的环境探索。核心机制在步骤2中通过松弛低相似度边使匹配图一致，步骤3中采用加权平均融合运动原语，并用诱导点集融合多个GP过渡。\n\n### 关键要点\n- SimFuse算法通过V2V/V2I通信实现多智能体间模型融合，支持同时融合n≥2个模型，突破了SILA只能两两融合的限制。\n- 采用余弦相似度计算运动原语间的相似度，综合考虑形状和位置重叠，用于确定哪些原语可匹配。\n- 引入匹配图与边松弛算法，通过移除低相似度边使每个连通分量内实现一对一的匹配，保证融合的一致性。\n- 融合运动原语时，对速度分量取元素级平均，对激活变量取并集，保留所有涉及的网格单元。\n- 过渡采用高斯过程建模，利用诱导输入（pseudo-inputs）将多个GP融合为新的稀疏GP，复杂度从O(N_d^3)降至O(N_I^2 N_d)。\n- 在五个公开非交叉数据集（ETH, Hotel, Univ, Zara1, Zara2）上，SimFuse的平均ADE为0.38，优于SILA的0.39及所有批量方法，FDE与SILA相当。\n- 在交叉路口场景（A、B、C、D、E）中，SimFuse平均ADE为0.88，优于Social-STGCNN的1.27和TASNSC的0.88，且融合后模型尺寸远小于不融合的SimFuse-naive。\n- 更新速度分析显示，SimFuse一次融合处理82条边仅需0.0033秒，而SILA融合24条边需0.001秒，但SILA需迭代34次，总耗时0.045秒，SimFuse总耗时0.0033秒，快约12倍。\n- 在分布式通信约束下（ad-hoc），使用对偶通信（每次仅与一个邻居）迭代融合，5个智能体经过8轮通信达到共识，预测误差逐步下降。\n- 相似性阈值β影响性能与模型规模，实验显示β=0.7在模型尺寸和ADE误差间取得最佳平衡。\n- SimFuse不依赖批量训练，能避免过拟合，融合多个次优模型可能产生比批量训练更好的准确率。\n- 实际应用场景包括自动驾驶中车辆间共享学习模型，无需传输原始数据，保护隐私同时降低通信负担。"
}