基于智能手机传感器的行为识别算法研究

发布时间：2018-12-25 07:49

【摘要】：随着科学技术与计算机技术的发展，无线网络的普及面越来越广，并且新兴科技下的产物可穿戴式传感器正在得到科研人员的认可与普及，故通过无线传感器发出信号，进行人体行为识别系统的开发已经成为具有重要研究意义与价值的工作，越来越多的科研机构开始利用行为识别系统进行广泛的科学研究。功能日益完善的智能手机也给人们的日常生活带来了极大的便利。随着在健康保健领域对行为识别系统的需求的增加，尤其是在老年护理，长期健康监控，以及协助有认知障碍患者，越来越多的注意集中在识别携带有传感器的人的行为上。智能手机上有很多软件能记录智能手机用户的日常行为。于是本研究在获取智能手机传感器信号的基础上，提出一种基于谱聚类和隐马尔可夫模型（Spectral clustering and Hidden Markov Models，SC-HMM）的日常行为识别算法。SC-HMM方法利用智能手机获取GPS地理位置、加速度、接收信号强度等传感器数据，结合谱聚类技术和隐马尔可夫模型学习，能有效地对用户日常活动行为进行自动识别。本研究通过在真实数据上进行实验验证本研究提出的SC-HMM识别方法。实验结果表明，在真实的智能手机数据集中，该方法具有较高的识别准确度，，并且优于以前传统的识别方法。本研究提出的识别方法在用户行为学习、情景感知等领域具有良好的实用性。本文的主要内容如下：（1）本论文研究无线传感器网络的架构，介绍了无线传感器的网络特征等方面的内容。（2）研究智能手机传感器技术、加速度传感器、RSSI技术，深入介绍了RSSI的原理， RSSI异常判断以及产生的原因等。研究如何收集智能手机传感器的传感器数据并分析智能手机用户的行为。（3）研究基于传感器的行为识别的分类，分为基于传感器的单人行为识别、基于传感器的多人行为识别；基于传感器的行为识别的识别阶段分为：在最底层，收集传感器数据，在中间阶段，采用统计推论，在最高层，识别出行为的目标或者子目标；基于传感器的行为识别的识别方法，主要有四种方法：概率推理方法、逻辑推理方法、基于WiFi的行为识别方法、基于数据挖掘的方法。本研究主要研究基于智能手机传感器的行为识别方法。（4）研究机器学习的聚类方法，谱聚类，并将谱聚类方法用于将从智能手机传感器中收集的传感器数据聚成K个相似的类。（5）研究无监督学习方法，隐马尔科夫模型的特性以及要解决的三个问题：估计问题、解码问题以及学习问题。通过隐马尔科夫模型训练活动时间序列得到智能手机用户的行为，并识别未训练的活动时间序列的行为。（6）总结本研究的工作，并展望未来的研究工作。可以将本研究提出的基于智能手机传感器的SC-HMM行为识别方法扩展到识别两个或者多个智能手机用户的交互行为识别。
[Abstract]:With the development of science and technology and computer technology, wireless network is becoming more and more popular, and wearable sensors, which are the product of new technology, are being recognized and popularized by researchers, so they send out signals through wireless sensors. The development of human behavior recognition system has become a work of great significance and value. More and more scientific research institutions begin to use behavior recognition system to carry out extensive scientific research. The increasingly sophisticated smartphone also brings great convenience to people's daily life. With the increasing demand for behavioral recognition systems in the field of health care, especially in geriatric care, long-term health monitoring, and assistance to patients with cognitive impairment, more and more attention has been focused on identifying the behaviour of people with sensors. There is a lot of software on smartphones that records the daily behavior of smartphone users. In this paper, based on the acquisition of smart phone sensor signals, a spectral clustering and hidden Markov model (Spectral clustering and Hidden Markov Models,) is proposed. SC-HMM). The SC-HMM method uses smart phone to acquire GPS data, such as location, acceleration and signal intensity, and combines spectral clustering technology with hidden Markov model to learn. Can effectively identify the daily activities of users automatically. The SC-HMM recognition method proposed in this study is verified by experiments on real data. The experimental results show that this method has high recognition accuracy in real smart phone data sets and is superior to the traditional recognition methods. The method proposed in this study has good practicability in user behavior learning and situational perception. The main contents of this paper are as follows: (1) this paper studies the architecture of wireless sensor networks and introduces the characteristics of wireless sensor networks. (2) the technology of smart phone sensor, acceleration sensor and RSSI technology are studied. The principle of RSSI, the abnormal judgment of RSSI and the cause of its occurrence are introduced in detail. This paper studies how to collect sensor data of smart phone sensor and analyze the behavior of smart phone user. (3) the classification of behavior recognition based on sensor is studied, which is divided into single person behavior recognition based on sensor and multi-person behavior recognition based on sensor. The recognition stage of behavior recognition based on sensor is divided into: at the lowest level, collecting sensor data, in the middle stage, using statistical inference, at the highest level, recognizing the target or sub-target of travel behavior; There are four main methods for behavior recognition based on sensor: probabilistic reasoning, logical reasoning, behavior recognition based on WiFi and data mining. This research mainly studies the behavior recognition method based on smart phone sensor. (4) the clustering method of machine learning, spectral clustering is studied, and the spectral clustering method is used to cluster the sensor data collected from smart phone sensors into K similar clusters. (5) study the unsupervised learning method, the characteristics of hidden Markov model and three problems to be solved: estimation problem, decoding problem and learning problem. The hidden Markov model is used to train the activity time series to obtain the behavior of the smartphone user and to recognize the behavior of the untrained activity time series. (6) summing up the work of this study and looking forward to the future research work. The proposed SC-HMM behavior recognition method based on smart phone sensor can be extended to identify the interaction behavior of two or more smartphone users.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP212.9;TN929.5

【参考文献】