面向多用户行为的无线识别关键技术研究

With the development and application of wireless communication technology, the wireless signals become ubiquitous. In addition to data communication, the changes of the wireless signals during the signal propagation also provide the ability of environmental perception. In particular, acquisition and analysis of human behaviors is a very important means in the association and integration of the information world and the physical world, which can be adopted widely in human-computer interaction, augmented reality and other prospects. Most of existing work however focus only on the action recognition, the study of activity, i.e., a sequence of action, and the detection of concurrent activities from multiple persons is poorly understood. This research strives to take advantage of the narrowband wireless signal to recognize the multi-person activities. To address the challenge that part of actions are easily confused, we establish a novel action correlation graph model to characterize whether or not two actions can be distinguished, explore the deployment policy in a 3D space to optimize the action recognition performance, and develop efficient machine learning algorithms to extract crucial signal features. The result is expected to enhance the capability of human activity recognition, expand the applications of wireless technology, and is therefore of great theoretical and practical value.

无线通信技术的发展与应用使得无线信号无处不在。除了用于数据通信，无线信号在传播过程中的变化也提供了环境感知的能力。对人类行为的获取与分析是实现信息世界、物理世界的关联与融合的重要手段，在人机交互、增强现实等应用前景广阔。现有工作多聚焦于动作的识别，对行为即动作序列及并发行为的研究甚少。本课题研究利用窄带无线信号实现多用户行为识别的关键技术，针对部分动作难以检测的问题，建立动作关联图模型描述动作的可区分性，探索三维空间的优化部署策略，研究基于机器学习的信号特征提取方法。研究成果将提升对人类行为的感知能力，拓宽无线技术的应用领域，具有重要的理论与实用价值。

无线通信技术的发展与应用使得无线信号无处不在。除了用于数据通信，无线信号在传播过程中的变化也提供了环境感知的能力。主要工作包括：(1) 驾驶行为分析：面向人机交互系统，以汽车驾驶为例，研究了基于无线信号分析的细粒度行为分析技术，它包括三方面的关键技术，一是基于深度卷积神经网络的质量特征提取；二是基于梯度的动作边界检测；三是基于行为融合的组合决策技术。(2) 手势识别系统：以手势识别为例，研究了基于WiFi信号的小尺度动作识别技术，提出了一套基于SVM的CSI信息分析方法，建立了CSI特征与手势的关联模型，通过实验评估了新方法的性能。(3) 影响因素分析：以手势识别为例，系统分析了部署策略、用户位置、通信机制、数据噪声等因素对无线识别系统的性能影响，总结了一系列实验规律，对基于机器学习的无线识别技术，阐明了下一步可能的发展方向。研究成果将提升对人类行为的感知能力，拓宽无线技术的应用领域，具有重要的理论与实用价值。