【作者】 肖玲；

【导师】 李仁发；

【作者基本信息】 湖南大学 ， 计算机应用技术， 2014， 博士

【摘要】 无线体域网是由可感知人体多种生理参数的轻便、可穿戴或可植入的传感器节点构建的无线网络。无线体域网为人体健康监测提供了新的手段,在疾病监控、健康恢复、特殊人群监护等领域有着巨大的应用意义和需求。通过佩戴在身体上的微惯性传感器,体域网可以采集人体的运动信号,在人体动作监测方面得到广泛应用,可实现人体动作识别、异常动作检测、步态识别与分析、运动能耗分析等目的。在利用无线体域网进行人体运动监测过程中,如何在满足身体活动监测指标要求的同时提高传感器节点的能量有效性,以便能在实际应用中长时间不间断地进行人体动作监测,是一个具有挑战性的问题。本文以由多个可穿戴的微惯性传感器构成的无线体感网为研究对象,围绕能量有效性,以稀疏表示和压缩感知理论为主线,从信号识别、信号压缩、数据融合、功率控制这四个方面展开研究。主要工作和创新点如下：(1)提出了一种基于自学习稀疏表示的动态手势识别方法L-SRC.针对手势识别中手势长短不一的问题,将手势样本向量进行归一化线性插值,从而将手势识别问题转化为求解待识别样本在训练样本中的稀疏表示问题；针对如何提高手势识别精度和速度的问题,采用基于类别的字典学习方法寻求一个较小的并经过优化的超完备字典来计算待识别样本的稀疏表示,从而在手势识别阶段大幅度缩减识别算法的计算复杂度,满足快速识别要求。在包含18种手势的数据集上验证了提出的L-SRC手势识别方法在保证识别精度的同时提升了识别速度。(2)提出了两种压缩分类的动作识别方法RP-CCall和RP-CCeach.针对运动信号的时间冗余性和稀疏性,结合压缩感知和稀疏表示理论,将传感信号压缩与动作识别相结合,以满足一定动作识别率的同时降低传感器节点的能耗。两种RP-CC方法是在传感器节点上利用随机投影对运动信号进行数字化的压缩采样,通过减少无线体域网的数据传输量来节省能耗；在基站上直接对压缩的数据建立稀疏表示的人体运动模式识别模型,利用稀疏系数的分布来实现动作识别。理论分析了压缩分类动作识别方法能正确识别的基本条件。找到了能在存储和计算资源有限的传感器节点上实现的随机投影矩阵。在包含13种动作的数据集上进行了验证,结果显示RP-CCall方法和RP-CCeach方法在对压缩的数据识别时也能达到无压缩时相近似的识别准确率,并高于最近邻、支持向量机等分类方法。(3)提出了基于分布式压缩感知和联合稀疏表示的动作识别方法DCS-JSRC.针对无线体域网中多传感器采集的运动数据之间的时空相关性,采用分布式压缩感知在传感器节点进行分布式压缩,充分利用这种相关性来进一步压缩数据以降低传输能耗。在基站通过探索多传感器节点感知运动信号的时空相关性,构建适用于动作识别的联合稀疏表示模型,将多传感器的动作识别问题转化为多变量稀疏线性回归问题来求解。采用层次贝叶斯模型来求解稀疏表示系数,利用不同传感器节点的相互关联来进一步提高动作识别的准确率。在动作数据集上进行验证,实验结果显示DCS-JSRC方法在相同压缩比的情况下获得了比RP-CCall方法和RP-CCeach方法更高的识别准确率。(4)设计了轻量级的基于动作行为的自适应功率反馈控制机制PID-A。针对无线体域网中链路通信质量受人的运动、姿态变化影响具有动态时变特性,通过实测人体不同动作以及发射功率变化对无线链路的影响,分析和总结了在人体不同运动状态下节点的发射功率与链路通信质量的变化特性和规律,在此基础上建立基于反馈的功率控制系统模型,结合人体动作识别的结果,来动态调整无线体域网中节点的发射功率。实验结果显示PID-A功率控制机制可保证在数据包成功接收的条件下降低了传感器节点发送数据包的平均能耗。(5)为了验证算法在实际系统中的性能,设计并实现了用于人体运动监测的无线体域网原型系统。利用所构建的基于微惯性传感器的无线体域网,采集人体在日常活动中的动作信号,实际验证了所提出的动作识别算法的识别准确率,并对传感器节点的能耗进行了分析,验证了算法的能量有效性。更多还原

【Abstract】 A Wireless Body Area/Sensor Network (WBAN/WBSN) is a kind of wireless network which is formed by physiological parameters sensors placed in the human body, on the body surface or around the body. As a pratical and innovative approach to improve health care and the quality of life, WBSN has been in great demand in remote monitoring of the physically or mentally disabled, the elderly, and children, medical diagnosis and treatment, physical rehabilitation and therapy. WBANs with inertial-based wearable sensors can collect motion signal of human body and have a broad range of applications in activity recognition, fall detection, estimation of energy expenditure, gait analysis and sports training.To improve the energy efficiency in meeting physical activity monitoring requirements is an important challenge for the practical deployment of WBSN in continuous long time of health monitoring. In this thesis, the WBSN with several wearable inertial sensors is taken as the research object, The energy efficiency, as the main concern of this thesis, is studied via the sparse representation and compressed sensing theory from4different aspects, whichi contain signal classification and signal compression, data confution, power control. The main contributions of the thesis are summarized below:(1) The L-SRC approach is proposed for accelerometer-based hand gesture recognition based on self-learning sparse representation. Linear interpolation is used on the acceleration signals to solve the problem of different lengths of traces for the same gesture so that the task of gesture recognition is casted as one of classifying among multiple linear regression models via sparse representation. Class-specific dictionaries learning is adopted and produced in an off-line procedure. The original sparse representation classification can be reduced from a large dictionary of all training samples to a small-sized one for sparse coding after dictionary learning, thereby obtaining significant speed-up. Experiments on the database of18hand gestures validate the performance of the proposed algorithm. The results show the L-SRC achieves high recognition accuracy while reducing the computing cost and time of recognition.(2)The RP-CCall and RP-CCeach mehods of compressed classification are presented for activity recognition. In view of the time redundancy and sparsity of the motion signal, these two approaches combine data classification with compression based on sparse representation and compressed sensing to reduce the energy consumption while maintaining the sufficient recognition accuracy of activity. The two approaches firstly compress the sensing data by random projection on the sensor nodes, and then recognize activities on compressed samples after transmitting to the central node by sparse representation, which can result in reduction of the energy consumption for data transmission. The performances of the two methods are evaluated on the Wearable Action Recognition Database (WARD) using inertial sensors placed on various locations on a human body. Experimental results show that the compressed classifiers achieve comparable recognition accuracies on the compressed sensing data. The recognition accuracies of our approaches are higer than the other classifiers such as the nearest nerbour and the support vector machine.(3)The DCS-JSRC approach is proposed, which explores a novel joint sparse representation method for activity classification of multi-sensor fusion. This approach utilizes the theory of distributed compressed sensing and joint sparse representation to develop a simultaneous dimension reduction and classification approach for multi-sensor activity recognition in BSNs. Both temporal and spatial correlations of sensing data among the multiple sensors are exploited for the purpose of compression and classification. Activity recognition with multiple sensors is formulated as a multi-task joint sparse representation model to combine the strength of multiple sensors for improving the classification accuracy. A hierarchical Bayesian modeling is used for simultaneous sparse approximation of multiple related signals. This method is validated on the WARD dataset. Experimental result shows that the DRP-JSRC achieves better classification performance than the RP-CCall and RP-CCeach·(4) The PID-A mechanism for dynamic power control of WBSN is proposed, which can adapt transmit power in real-time based on feedback information in order to simultaneously satisfy the requirements of energy efficiency and link reliability. The properties of the link states by using the recived signal strength indicator under different scenarios with body posture change and dynamic body motions are first experimentally investigated. The empirical evidence of real human environment shows the rapid change of wireless link quality in WBSN. The dynamic nature of on-body links with varing body postures is characterized based on experimental results. The PID-A mechianism utilizes the feedback results of activity recognition and RSSI to adjust the optimal transmission power assignments. The performance of the PID-A is experimentally evaluated and compared with a number of static power assignment schemes. Experimental results show that the PID-A can gain higher packet delivery rate and reduce the energy consumption of a packet at the same time.(5) A prototype of WBAN is designed and developed for verifying the performance of our proposed approaches. The real gesture and activity data were collected by using our prototype platform. The proposed mehods are verified on real dataset and the recognition accuracies of our methods are derived. Energy consumption is measured and analsized based on the energy profile of node platform and the energy efficiency of our proposed approaches is demonstrated.更多还原