WiSPE: A COTS Wi-Fi-Based 2-D Static Human Pose Estimation
WiSPE: A COTS Wi-Fi-Based 2-D Static Human Pose Estimation
复制标题
DOI：
10.1109/jsyst.2023.3270495
复制
发表时间：
2023-09
影响因子：
4.4
通讯作者：
Mingming Xu;Zhengxin Guo;Linqing Gui;Biyun Sheng;Fu Xiao
中科院分区：
文献类型：
--
作者：
Mingming Xu;Zhengxin Guo;Linqing Gui;Biyun Sheng;Fu Xiao
随着商用现货 (COTS) 设备的广泛部署，基于 Wi-Fi 的人体姿态估计因其无处不在的可用性而在无线传感领域引起了广泛关注。由于人类活动传感是以人为本，人体姿态估计已成为重要的基础设施之一。在本文中，我们提出了一种基于 Wi-Fi 的二维静态人体姿势估计（WiSPE），它就像相机一样来估计骨骼中的人体姿势。具体来说，在设备部署中，我们设计了带有接收天线的垂直滑轨，以获取不同高度的信道状态信息（CSI）。然后，我们利用多信号分类算法从 CSI 计算到达角 (AoA) 和飞行时间 (ToF) 频谱。我们处理不同高度的 AoA 和 ToF 频谱，并构建一种称为 2-D AoA 图像的新颖特征，该特征捕获由人体引起的多路径信号反射。此外，我们提出了环境背景滤波器（Env-Filter）算法，从静态环境信号中创建一种与环境相关的噪声滤波器。 Env-Filter算法有效滤除环境背景噪声，增强二维AoA图像中的人体姿态相关信息。最后，我们设计了一个师生网络来将骨骼与二维 AoA 图像相关联。实验结果表明，WiSPE的平均正确关键点百分比（PCK）@50的准确率高达95.2%，达到了与基于视觉的方法相同的效果。当评估标准变得更严格到PCK@10时，WiSPE的平均准确率仍然达到73.1%，这优于最先进的基于Wi-Fi的方法。
With the wide deployment of commodity off-the-shelf (COTS) devices, Wi-Fi-based human pose estimation has attracted significant attention in the field of wireless sensing for its ubiquitous availability. Since human activity sensing is human oriented, human pose estimation has become one important infrastructure. In this article, we propose a Wi-Fi-based 2-D static human pose estimation (WiSPE), which acts like a camera to estimate human pose in skeletons. Specifically, in the device deployment, we design a vertical slide rail with receiving antennas to obtain channel state information (CSI) at different heights. Then, we utilize the multiple signal classification algorithm to compute angle of arrival (AoA) and time of flight (ToF) spectrum from the CSI. We process the AoA and ToF spectrums at different heights and construct one novel feature called 2-D AoA image, which captures multipath signal reflections caused by the human body. Furthermore, we propose the environment background filter (Env-Filter) algorithm to create one environment related noise filter from the static environment signals. The Env-Filter algorithm effectively filters out the environment background noise and enhances the human-pose-related information in the 2-D AoA image. Finally, we design a teacher–student Network to correlate skeletons with the 2-D AoA image. The experiment results show that the average accuracy of the WiSPE is up to 95.2% in percentage of correct keypoints (PCK)@50, achieving the same effect as vision-based methods. When the evaluation criterion gets stricter to PCK@10, the average accuracy of the WiSPE still achieves 73.1%, which is better than state-of-the-art Wi-Fi-based methods.