面向活动识别的多源多维传感器数据融合、交互和依赖问题的数学理论和方法研究

项目名称： 面向活动识别的多源多维传感器数据融合、交互和依赖问题的数学理论和方法研究

项目编号： No.61473339

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 自动化技术、计算机技术

项目作者： 王金甲

作者单位： 燕山大学

项目金额： 58万元

中文摘要： 从嵌入式或穿戴式传感器数据中自动识别人类活动是普适计算的一个关键和富有挑战性的研究课题。目前复杂活动识别仍然是一个开放的研究挑战，需要多学科的努力，其正确率低的主要原因是：过去的传感器数据融合方法，限制了识别正确率，没有完全考虑不同传感器间的交互对分类结果的影响，将距离很远时间的数据远程依赖关系进行合并很困难。因此主要研究内容包括：针对特征层融合问题，研究稀疏组套索的传感器选择、特征选择和分类模型及坐标下降法；针对决策层融合问题，研究分量套索模型，具体包括分量的选择、建模和合并；针对数据交互问题，研究分层套索和分层组套索模型及坐标下降等凸优化算法。针对数据远程依赖问题，研究广义图套索模型及凸优化算法；研究构建在线活动识别软硬件平台。将套索罚分类模型及凸优化算法的最新统计学研究成果应用于信息领域的复杂活动识别问题，是一定能做出特色性研究成果的。本项目对其它多传感器数据融合问题是很好的借鉴。

中文关键词： 活动识别；多传感器信息融合；超高维数据；套索；凸优化

英文摘要： Recognising human activities from sensors embedded in an environment or worn on bodies is an important and challenging research topic in pervasive computing. Now,complex activity recognition remains an open research challenge and requires a multi-disciplinary effort. The main reasons of the low accuracy have there points. The past sensor data fusion methods limit the recognition accuracy of classification results, not fully consider the influence of interaction among different sensors on classification results and have the difficulty of incorporating long-range dependencies between distant time instants. So the main research contents include: for feature layer fusion problem, the research of sensor selection, feature selection and classification model via sparse group lasso and coordinate descent; for decision layer fusion problem, the research of component lasso model including component selection, component modeling and component combination; for data interaction problem, the research of hierarchical lasso model, hierarchical group lasso model, and coordinate descent convex optimization algorithms; for data dependence problem, the research of generalized graphical lasso model, and convex optimization algrithms; the research of the construction of software and hardware platform for online activity recognition using. The new statistical research results of the lasso penalized classification model and convex optimization algorithm are applied to the complex activity recognition problems of information field. It is sure to make the characteristic research. This project is a good reference for other multiple sensor data fusion problem.

英文关键词： activity recognition;multi-sensor information fusion;Ultra high dimensional data;lasso;convex optimization