基于多层次特征分析的复杂人体活动识别方法研究

Accurately and robustly recognizing human activities plays its basic role in various pervasive applications such as smart elderly healthcare and human-computer interaction. Approaches based on wearable sensors have gained substantial attention due to their pervasiveness, all-day for use, and low cost. However, most of existing solutions suffer from limited performance due to insufficient information utilization, lack of multi-channel cooperative analysis, and confused similar activities. After modelling the raw sensor signals with multi-channel multi-level features, this proposal is to launch a systematic and deep study on several issues to be tackled for the complex human activity recognition setting. More specifically, this study focuses on the following problems: (1) representation of sensor signals with multi-level informative features and the optimal feature space search; (2) cooperative analysis of multi-channel streaming signals and cross-modal learning for heterogeneous sensors; (3) data-driven hierarchical fine-grained activity recognition model and its incremental update method. The objective is to provide the wearable human activity recognition theory model and its high accuracy, high reliability method. Our research will surely make an important advance in the application of complex activity recognition and offer corresponding theoretical and technological support.

在智慧健康养老及人机交互等应用与研究领域中，准确鲁棒地识别人体活动具有重要的基础性作用。基于可穿戴传感器的方法因其普适性、全天候性及成本较低受到了高度重视，但是现有的解决方案存在信息利用不充分、多信道协同分析不足、相似活动易混淆等问题。本项目致力于研究复杂人体活动识别中的关键技术，提出通过多信道多层次特征来实现对原始传感器信号的建模和分析，并围绕技术路线中若干有待解决的问题开展研究。主要研究内容包括：（1）传感器信号的多层次信息特征表示和特征空间优化；（2）多信道传感器信号的多层次非线性协同优化分析和跨模态特征学习；（3）建立基于数据驱动的层次化细粒度动态活动识别模型。本项目以提供可穿戴计算情境下高准确率、高可靠的复杂人体活动识别模型和方法为目标，为活动识别技术的应用提供理论基础与核心算法支撑，具有显著的理论意义和应用价值。

在智慧健康养老及人机交互等应用与研究领域中，准确鲁棒地识别人体活动具有重要的基础性作用。基于可穿戴传感器的方法因其普适性、全天候性及成本较低受到了高度重视，但是现有的解决方案存在信息利用不充分、多信道协同分析不足、相似活动易混淆等问题。本项目深入研究复杂人体活动识别中的关键技术，开展如下研究工作：研究传感器信号的特征表示和特征空间优化，分析不同特征编码模式的信息特征表示能力以及与不同活动识别模型的组合方式；研究传感器信号的多层次非线性协同分析，建立基于多视图融合的活动识别模型；研究人体活动之间的语义层次性，建立数据驱动的层次化活动识别模型。研究成果包括发表论文11篇，其中在Expert Systems with Applications、IEEE Sensors Journal等国际SCI期刊上发表论文6篇，授权发明专利1项、登记软件著作权2项，获得安徽省电子信息科技奖科技进步一等奖奖励1项，参与CCML、BIBM等国内外学术交流会议十余次，培养硕士生4名。项目执行情况良好，完成项目所有目标。