基于商用设备的低代价高精度定位与姿势识别研究

Localization and posture recognition based on wireless signals have advantages of low light condition requirement and privacy security, thus have many potential applications, such as security monitoring, elderly people guardianship and computer-human interaction, etc. However, existing dedicated devices based methods have a high implementation cost, training methods have problems of high deployment cost, low accuracy, sensitive to noise and subjecting to multipath. To solve these problems, this project explores the use of commercial wireless devices to implement low-cost, high accuracy localization and posture recognition methods. The research contents include: the high accuracy localization method on the framework of compressed sensing with multipath profile based robust fingerprinting features designing, the high robust posture recognition method based on the signal mapping and transformation which removes the influence of position and orientation diversity, the multipath and noise digestion method based on signal calibrating and “pure” subcarrier selection, reducing the impact of multipath on localization and recognition accuracy, improving the environmental adaptability. It’s scientific essence is to explore the constraint relationship between deployment cost and accuracy; explore the interference pattern on wireless signal from position, direction, and user postures; and explore the feasibility and ubiquitous of low cost wireless sensing. The results will provide a theoretical basis for human-computer interaction and other applications, and have great significance to promote the Internet of Things.

由于基于无线信号的定位和姿势识别具有对光线等条件要求低与隐私安全等优势，在安防监控、老人监护和人机交互等领域前景广泛。然而现有基于专用设备的方法存在着实施成本高，基于学习的方法存在着部署代价高、精度低、噪声敏感、受制于信号多径等问题。对此，本项目探索利用普适商用无线设备实现低成本、高精度定位与姿势识别的有效方案。具体研究内容包括，利用基于无线信号多径轮廓的高区分度指纹特征并结合压缩感知的高精度定位方法；基于映射变换消除位置-方向多样性的高鲁棒目标姿势识别方法；基于无线信号校准和“纯净”子载波的多径与环境噪声消解方法，降低多径对定位与识别精度影响、提高环境适应性。其科学实质是寻求部署代价与定位及识别精度间的约束关系，发现目标位置、方向与姿势对无线信号影响的规律，探索低成本无线感知技术的可行性和普适性。研究成果将为无线感知和人机交互等应用提供理论依据与技术参考，对推动物联网应用有重要意义。

本项目针对现有基于指纹的姿势识别系统在识别目标对象改变时，为保证高精度需要收集所有测试目标的训练数据从而耗费大量时间和人力成本的问题，提出了低代价高精度的跨目标姿势识别方法。通过收集小样本的新用户数据，设计特征提取算法挖掘只与姿势有关但与用户身份无关的特征，利用神经网络对抗模型合成大量的训练数据，实现用户自适应的识别模型；进一步地，通过分析发现数据生成模型的鲁棒性低，提出基于视频数据生成姿势Wi-Fi数据的模型，并考虑真实环境中多径干扰的影响，引入Unet架构建立生成数据与真实测量数据的特征映射关系，实现高精度的跨目标手势识别。另一方面，项目组针对物联网新兴应用需求，提出基于Wi-Fi的反向散射标签定位和信息传输系统，深入分析标签信号传输模型，发现多标签混合信号可以等效为单标签多路径叠加的信号，提出多路径分离算法提取标签有用信号剥离多径信号的影响，建立信道参数与时间相关的统计特征并与传输信息建立关联映射，同时实现定位和传输的功能。此外，项目组提出基于无线信号的材质识别方法，通过构建材质与信号的映射关系提取材质识别的细粒度特征，实现高精度的材质识别系统。本项目的相关研究成果陆续发表在CCF推荐A类会议UbiComp，B类会议ICDCS，A类期刊TMC上，授权技术发明专利1项。姿势识别和定位、材质识别是物联网应用的重要组成部分，本项目的研究对物联网在实际应用中的发展具有重要的意义。