基于集成学习的多传感器选择和在线人体行为识别研究

With the rapid development of smart home and wearable device fields, the applications of multi-sensor based online activity recognition are increasingly widespread. Therefore, to develop and improve the multi-sensor based online activity recognition system is essential and significant. However, current multi-sensor activity recognition system suffers from the problems of lacking of uniform sensor layout, unstable performance of online learning, which are emergency and critical to be solved for application area extension. This project aims to research on adaptive multi-sensor selection problem in activity recognition deployment and analyze multiple classifiers based online learning problem in activity recognition system. In the framework of ensemble learning, by combining online learning and ensemble diversity methods, the project will study the effective recognition based on real-time data stream with concept evolution and imbalance class characteristics. Further, two adaptive strategies and ensemble selection approaches will be designed to optimize the multi-sensor selection. Overall, the project will investigate the theory basis for generalized multi-sensor activity recognition and achieve the theoretical innovation and technological breakthroughs for multi-sensor selection， which will promote the research and development of the activity recognition related research fields.

随着智能家居和可穿戴设备领域的迅速发展，通过配置多个传感器以进行在线识别人体行为的应用日益广泛。因此，对基于多传感器的在线行为识别系统的开发和完善具有重要的研究意义。然而，目前多传感器行为识别系统仍存在设施布局不统一、在线学习识别性能不稳定等推广应用急需解决的关键性难点问题。本项目旨在研究人体行为识别配置中的传感器自适应选择问题，模型化分析人体行为识别中多分类器的在线学习问题，在集成学习框架下，结合在线学习和集成差异度等方法以研究针对具有概念演变和类别不平衡特点的人体行为实时数据流的高效识别，并在此基础上研究自适应策略和分类器选择方法，最终实现多传感器的优化配置。本项目探索具有普遍性的多传感器人体行为识别理论，实现多传感器选择的理论创新与技术突破，促进人体行为识别领域的研究与发展。

本项目针对基于多传感器的人体行为识别问题，利用机器学习的集成学习等理论，经过三年的研究，在可穿戴多传感器选择，在线概念演变检测、特征提取、行为标注、状态识别等方面的理论与方法开展了深入的研究。通过本项目的研究，提出了基于集成剪枝的多传感器自适应优化选择人体行为识别系统，保证可穿戴场景下用户的个性化和准确度的平衡性；研究了基于拒绝规则的在线学习优化人体行为识别问题，减轻了概念演变的负面影响；设计了基于深度学习的特征提取算法，提高了人体行为识别系统性能；拓展了人体行为自适应和状态识别系统，有效地应用到驾驶和轮椅等实际场景中。本项目也对深度学习等当前计算机领域的热点问题和方法展开了相关的前期探索，研究成果具有较强通用性，为进一步研究和应用到基于传感器的人体行为识别问题奠定了坚实的基础。本项目发表论文10篇，其中SCI检索3篇，EI检索6篇，软著授权1个，培养博士研究生2名，硕士研究生4名，本科生4名。