基于智慧家居感知数据的老人日常行为识别与异常检测

发布时间：2018-04-08 14:41

本文选题：智慧家居　切入点：老人日常行为识别　出处：《杭州电子科技大学》2017年硕士论文

【摘要】：随着人口老龄化情况加剧、人力资源短缺,养老成为当前社会面临的主要问题。家庭环境中自动识别老人日常行为,发现老人行为异常,提高老人独立生活能力和家庭健康护理水平,是缓解养老困境的可行方法。本文以华盛顿州立大学智能空间实验室的实验数据为基础,研究机器学习理论、数据处理技术,结合智慧家居相关技术,识别老人日常行为,了解老人的意图,发现老人生活中的行为异常与环境异常。主要研究内容如下所示:(1)设计传感器事件动态分割与标注方法。根据活动执行过程中所触发传感器之间的依赖关系,计算相邻窗口的相似度,过滤行为边界点,保留相似度最低的点,最后合并相邻的且具有相同标签的序列片段。(2)构建以较高精度识别老人多种日常行为的模型。根据老人日常行为特点,构建多维度特征框架,包括传感器权重、传感器类别特征、时间特征和频率特征;使用遗传算法和交叉验证对行为模型进行训练与调优,建立基于支持向量机的日常行为模型。(3)设计老人异常行为检测框架。结合日常行为模型,选择重要特征,建立老人各类行为的高斯混合模型,计算异常样本与所属成分的偏离程度,对比各类特征与正常样本特征之间的差异。日常行为识别是实现老人生活辅助、提升老人独立生活能力、提高健康护理水平的前提。设计多组实验评估本文方法和模型的性能。实验结果表明序列分割方法能够对传感器序列进行合理分割,日常行为模型能准确的识别多种行为,异常检测方法能够高效地发现老人行为中的异常。研究结果对居家养老智慧化发展具有一定促进作用。
[Abstract]:With the aggravation of the aging population and the shortage of human resources, old-age support has become the main problem facing the society.It is a feasible method to identify the daily behavior of the elderly automatically in the family environment, find out the abnormal behavior of the elderly, improve the independent living ability of the elderly and the level of family health care, which is a feasible method to alleviate the plight of the aged.Based on the experimental data from the Intelligent Space Laboratory of Washington State University, this paper studies the theory of machine learning, data processing technology, combining with smart home related technology, to identify the daily behavior of the elderly and to understand the intention of the elderly.To find out the abnormal behavior and environment in the old people's life.The main contents of this paper are as follows: (1) Design the sensor event dynamic segmentation and tagging method.According to the dependencies between sensors triggered during the execution of the activity, the similarity of adjacent windows is calculated, the behavior boundary points are filtered, and the points with the lowest similarity are retained.Finally, the adjacent sequence fragment with the same tag is merged to construct a model to identify various daily behaviors of the elderly with high accuracy.According to the characteristics of daily behavior of the elderly, a multi-dimensional feature framework is constructed, including sensor weight, sensor category feature, time feature and frequency feature, and the training and tuning of the behavior model using genetic algorithm and cross-validation.A support vector machine (SVM) based daily behavior model is established to design a framework for the detection of abnormal behaviors of the elderly.Combined with the daily behavior model, the Gao Si mixed model of all kinds of behaviors of the elderly was established, and the deviations between abnormal samples and their components were calculated, and the differences between different characteristics and normal sample characteristics were compared.Daily behavior recognition is the premise to realize the elderly's life assistance, improve the elderly's independent living ability and raise the level of health care.Several experiments were designed to evaluate the performance of the proposed method and model.The experimental results show that the sequence segmentation method can segment the sensor sequences reasonably, the daily behavior model can accurately identify a variety of behaviors, and the anomaly detection method can efficiently detect the anomalies in the behavior of the elderly.The results of the study have a certain role in promoting the development of home-based endowment intelligence.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TU855;TP181

【参考文献】