{
  "paper_id": "ef4ae11082d89f991f7f3cfbe51d69b5",
  "paper_title": "Privacy-Preserving User Pose Prediction for Safe and Efficient Human-Robot Interaction",
  "paper_authors": [
    "Adel Baselizadeh",
    "Weria Khaksar",
    "Md Zia Uddin",
    "Diana Saplacan",
    "Jim Torresen"
  ],
  "publication_date": "",
  "publication_venue": "2023 IEEE 19th International Conference on Automation Science and Engineering (CASE)",
  "abstract": "Enhancing user privacy is crucial in improving the safety and efficiency of Human-Robot Interaction (HRI), as it is a key factor for establishing user trust in the robot. Using privacy-preserving sensors and local processing of the user's data are ways to enhance privacy in HRI. This paper presents a privacy-preserving sensing system for real-time tracking and predicting the user's movements in HRI. As privacy-preserving sensors, a thermal and a depth camera are used to monitor the user's movements and determine their current pose. In order to improve the robot's perception of the user's situation and enhance the quality of real-time user monitoring, a Deep Learning (DL) model has been developed to estimate the future poses of the user. The developed model is based on the Sequence to Sequence mechanism (Seq2Seq). Modifications have been made to Seq2Seq so it can be run locally on the robot. As a result of these modifications, the computational cost of the model has been reduced by 34%. Experimental studies have been conducted to evaluate the performance of the sensing system in tracking and predicting the user's movements in HRI. According to the test results, the proposed sensing system is able to track the movements of the user appropriately. Additionally, it is shown that the estimation of the user's movement through the proposed system with the prediction model can improve the safety and efficiency aspects of the HRI experiments by up to 24% and 17%, respectively.",
  "keywords": [
    "Privacy-preserving sensing",
    "User pose prediction",
    "Human-Robot Interaction",
    "Thermal camera",
    "Depth sensor",
    "Seq2Seq",
    "Deep learning"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:13.282655Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文针对人机交互（HRI）中的用户隐私保护问题，提出了一种隐私保护的用户姿态感知与预测系统。该系统使用热像仪和深度相机作为隐私保护传感器，通过Openpose处理热图像获取用户2D关节点，再通过多传感器标定（Homography矩阵）与深度数据融合，得到用户上肢8个关节的3D位置，从而实现实时用户监测。为提升机器人对用户未来状态的感知并补偿传感器处理延迟，文中开发了基于Seq2Seq的轻量化姿态预测模型，通过同时引入残差连接、注意力机制、双向RNN和Time2Vec嵌入等五种修改，在保持精度不变的同时将计算成本降低34%。模型使用INHARD数据集（工业HRI场景中16位受试者共13种任务的数据）训练，输入过去2秒的关节3D位置序列，预测未来1秒的位姿。实验采用TIAGo机器人进行手腕跟随测试和人机碰撞测试，结果表明预测模型可将轨迹跟随精度（RMSE）在中速和快速下分别提高约25%和10%，将机器人碰撞反应时间缩短最多约15%，并可将机器人退出危险区后的反应时间提前最多约17%，从而显著提升了HRI的安全性和效率。未来工作将探索更复杂的协作场景、使用AlphaPose等替代库以及非视觉传感器的融合。\n\n### 关键要点\n- 提出一种结合热像仪和深度相机的隐私保护感知系统，用于实时追踪用户上肢姿态，取代RGB相机以增强隐私。\n- 利用Openpose库对热图像进行实时2D姿态估计，并采用Homography矩阵及伪逆算法完成热像仪与深度传感器的标定融合，得到8个上肢关节的3D位置。\n- 开发的姿态预测模型基于Seq2Seq架构，并同时应用残差连接、Bahdanau注意力、双向RNN、Time2Vec嵌入和波束搜索（后因耗时被排除）五种修改，实现本地轻量化运行。\n- 修改后的Seq2Seq模型在保持85%预测精度的同时，计算成本较原始Seq2Seq降低34%，运行时间最短。\n- 模型训练使用INHARD工业HRI数据集，以15 FPS采样，输入2秒历史关节位置，输出未来1秒预测，采用Adam优化器，损失函数为均方误差。\n- 硬件配置包括Optris P1450热像仪和Orbbec Astra S深度传感器，Openpose在NVIDIA Quadro RTX 3000 GPU上运行，关节数据获取速率约15 FPS，比RGB处理慢约20%。\n- 手腕跟随实验中，预测模型可提升轨迹跟随精度：中速整体RMSE改善24.97%，快速改善9.83%，但在快速Y方向出现7.76%的负提升，表明预测模型在快速运动下有局限性。\n- 人机碰撞测试中，预测用户手部位置可使机器人避险反应时间缩短最多15.03%（XY平面），手离开危险区后机器人重新开始任务的反应时间最多提高17.58%（YZ平面）。\n- 所有实验通过UR5机器人控制用户手部运动以保证轨迹一致，测试重复多次并覆盖不同方向以增强可靠性。\n- 预测模型不仅提升安全性和效率，还能补偿热像仪图像处理带来的系统滞后，但快速运动时预测精度下降影响性能。\n- 未来工作建议使用更丰富的训练数据集、AlphaPose等其他关节点检测库，以及融合非视觉传感器以进一步增强隐私感知能力。"
}