基于条件随机场的参与式跌倒检测方法

发布时间：2018-05-14 01:22

本文选题：参与式感知 + 特征提取　；参考：《燕山大学》2016年硕士论文

【摘要】：我国人口老龄化问题日益严重,老年人的健康安全监护已经成为移动计算技术研究热点之一。跌倒作为威胁老年人的身心健康的重要因素之一,受到学术界和工业界研究人员的普遍重视。现有的跌倒检测系统局限于计算复杂度高、需要外部部署以及可扩展性差等诸多因素,使其不能得到广泛应用。论文基于参与式统计感知方法,在人体动作特征识别分类和利用条件随机场模型判定跌倒两面方面做了深入研究。首先,提出基于统计感知的动作识别分类的方法。利用智能手机内置传感器收集人体动作信号,进行过滤降噪后,提取峰值特征、协方差特征和频域特征,利用支持向量机将人体动作分为日常动作、跌倒及疑似跌倒动作、步态异常和静止四大类,为后续上下文识别跌倒行为提供标注依据。其次,提出基于条件随机场的跌倒检测模型。根据人体动作的分类和动作之间的上下文联系,进行动作标注和特征标注;通过极大似然估计和梯度算法进行条件随机场模型参数训练,然后在给定条件随机场模型参数的情况下,利用维特比算法预测出最可能的动作序列。最后,使用多种智能手机采集传感器数据,在Matlab平台和Visual Studio平台上进行实验,对本文提出的特征提取、行为分类和跌倒检测方法进行全面的性能分析、阈值确定和准确度分析。
[Abstract]:The problem of aging population in China is becoming more and more serious, and the health and safety monitoring of the elderly has become one of the hotspots of mobile computing technology. Falling, as one of the important factors threatening the physical and mental health of the elderly, has been paid much attention by researchers in academia and industry. The existing fall detection system is limited to high computational complexity, requires external deployment and poor scalability, so it can not be widely used. Based on the participatory statistical perception method, this paper makes an in-depth study on the recognition and classification of human motion features and the use of conditional random field model to determine the two sides of the fall. Firstly, an action recognition classification method based on statistical perception is proposed. Using smart phone built-in sensor to collect human action signal, filter noise, extract peak feature, covariance feature and frequency domain feature, use support vector machine to divide human action into daily action, fall and suspected fall action. Gait anomaly and stillness provide the basis for subsequent context recognition of tumble behavior. Secondly, a fall detection model based on conditional random field is proposed. According to the classification of human actions and the context relation between actions, the action tagging and feature tagging are carried out, and the conditional random field model parameters are trained by maximum likelihood estimation and gradient algorithm. Then the most probable action sequence is predicted by using the Viterbi algorithm under the condition of random field model parameters. Finally, we use a variety of smart phones to collect sensor data and carry out experiments on Matlab and Visual Studio platforms. The performance analysis, threshold determination and accuracy analysis of the proposed feature extraction, behavior classification and fall detection methods are carried out.
【学位授予单位】：燕山大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP274

【相似文献】