面向信息物理系统的人体行为识别理论与关键技术研究

Cyber-Physical Systems, CPS is the combination of physical and digital world. It has been gained widely attentions overall the world. The core goal of the CPS is to pursue a highly efficient, lightweight, ubiquitous development. Driven by this goal, the human activity recognition using commercial off-the-shelf Wi-Fi devices appears. The new technique is very competitive with its unique characters. However, there are still lots of challenges. Firstly, the Channel State Information (CSI) is not completely extracted, which influences the accurate recognition. Secondly, the existing research mostly depends on the training, which introduces extra overheads. Thirdly, the scenes and activities constraints limit the scope of the applications. To tackle these issues, two underpinnings will be proposed: a CSI-Doppler Shift model and Doppler Shift-activity direction model. The fully Doppler Shift information will be extracted by the first model and the second model will be adopted for the accurate activity direction recognition. An activity recognition framework will be developed with the distinct features of light-weight, training free, no constraints and ubiquity. This will provide CPS oriented technology with theoretical and practical supports.

由物理世界与数字世界相互融合构成了物理信息系统（Cyber-Physical Systems，CPS）已经受各国广泛关注。在CPS的高效率、低成本、轻量级、普适化的发展目标驱动下，应运而生的是基于商用无线Wi-Fi人体行为感知与识别技术。虽然该项技术以其独特优势，在竞争中占据不败之地，但是仍面临诸多挑战。首先，信道状态信息未曾有效完整提取，影响了识别效果。其次，前期的工作多基于繁琐的训练，造成大量额外开销。最后，对场景和人的行为的约束，限制了方法的普适性。针对这些问题，本课题拟研究面向物理信息系统的人体行为感知识别理论与技术。通过建立信道状态信息与多普勒频偏的关系模型以及多普勒频偏与人体行为方向的关系模型，实现对多普勒频偏的完整、有效地提取，摆脱繁琐的训练，实现轻量级、无约束的精准行为方向识别。为最终形成统一的人体行为识别框架以及面向CPS的行为感知与识别提供理论基础和技术支持。

本项目在CPS技术的总体目标驱动下，研发以人为本的、轻量级、普适性好的智能感知识别系统，为无线人体感知理论和技术的发展提供必要的技术支持。项目围绕该领域的关键问题：多人同时存在的行为监测、连续复杂动作的分割、环境改变导致的行为识别成功率的下降展开研究。建立CSI动态变化和人的各种行为之间的模型。提出了基于迁移学习的消除环境依赖的跨域行为识别方法；基于张量分解的多人重复性运动监测；基于菲尼尔区理论的消除路径、方向依赖的步态识别；基于进出信号变化差异的人群移动方向估计，并构建基于支持向量机（SVM）改进模型的分类算法来实现跨环境人群计数；将DBN和HMM与高斯混合相结合实现美国手语连续句子的识别；基于手动特征提取与自动特征提取相结合的暴力行为识别。针对具体应用场景特点，以人为核心，选择模型和模式的算法，在保证识别精度的前提下，追求算法的低复杂度、高普适性、高可行性、强鲁棒性、无约束性。成果发表在该领域的多个顶级期刊和会议上。