当前位置: 首页>博士论文>资源详情
基于可穿戴传感器数据的人体动作识别方法的研究
中文摘要

人体动作识别已经受到了广泛的关注,它在医疗、安全、娱乐和军事等许多领域都具有重要应用。传统的基于视频的动作识别方法,容易受到物体遮挡、光线条件等多种因素的制约,且不利于用户隐私的保护。随着微电子和无线通讯技术的快速发展,基于可穿戴传感器的动作识别已成为研究热点。与基于视频的方法相比较,传感器的高计算能力、小尺寸和低成本等特点,使得人们可以将这些设备作为日常生活的一部分与其进行交互。这将极大地促进移动健康管理以及运动监测等多种应用的发展,在普适计算中具有重要意义。本文使用可穿戴传感器节点,采集并存储运动时的加速度和角速度数据,对人体动作识别的方法进行了深入分析,具体工作包括: 1.采用单个传感器节点对人体下肢动作进行了识别。针对动作识别实时性的需求,提高分类效率是需要解决的关键问题。为此,本文提出一种基于模糊聚类的方法,将大量的训练样本归为少数聚类中心,并基于这些中心向量和隶属度函数设计分类器。实验结果表明,相比于一些常见的分类算法,该方法获得了较高的识别正确率。而且,它是基于小样本的数据集识别下肢动作,不但缩短了分类器执行任务所需的时间,同时也降低了系统的存储需求。 2.目前已有的动作识别研究主要针对一些简单的基本动作进行分类,对于上肢和下肢同时产生运动的情形缺少分析。为此,本文提出一种分层机制和神经网络相结合的方法,将复杂的并发动作识别分为两个阶段的任务。其中,底层采用一个神经网络对人体的下肢动作进行分类,而顶层则采用多个神经网络建模上肢动作,并推断出最终的并发动作。本文提出的方法将整体的分类问题分解成两个子问题,降低了决策边界的复杂性,识别效果明显优于基于单层的方法。 3.在复杂的动作识别中,需要充分挖掘动作时间序列数据间的关联。为此,本文提出了一种基于平行隐马尔可夫模型的方法,采用一组具有时间顺序的特征向量序列来描述动作的持续和转移。提取统计时域和频域特征作为观测向量,并使用主成分分析的方法进行特征降维。在动作的识别阶段,平行隐马尔可夫模型结合两个信道的概率做出最终的决策。实验结果表明,本文提出的基于序列的识别方法,在性能上优于基于单示例的一些分类器。 4.基于可穿戴传感器的人体动作识别研究多集中于讨论单人动作,很少有工作分析双人交互动作。本文针对日常生活中常见的交互动作,采用动态时间规整和马尔可夫逻辑网相结合的方法对其进行分类。其中,动态时间规整算法用来分类单人动作,马尔可夫逻辑网则通过语义建模双人交互动作。马尔可夫逻辑网为每个规则分配了一个相关联的权重,使得它在进行决策时具备一定的纠错能力,可以帮助提升双人交互动作识别的正确率。 关键词:可穿戴传感器;人体动作识别;识别正确率;特征提取;主成分分析

英文摘要

The study of human activity recognition has received wide attention, which may play an important role in many ways including medical, security, entertainment and military scenarios. The traditional video-based methods may be constrained due to object occlusion and light conditions, and they are not conductive to user privacy protection. With the development of microelectronics and wireless communication technology, the sensor-based recognition methods have become a research hotspot. Compared with the video-based methods, wearable sensors have high computational power, small size and low cost, allowing people to interact with the devices as part of their daily living. This will greatly motivate the development of many applications such as mobile health management and motion monitoring, and has important significance in pervasive computing. In this study, acceleration and angular velocity data are collected and stored by using wearable sensor nodes. The main work consists of the following parts: 1.Only one wearable sensor node is used to recognize human lower limb activity. For the real-time requirement of activity recognition, how to improve the classification efficiency is a key problem to be solved. In this study, a fuzzy clustering algorithm is proposed to convert lots of training samples into a small number of clusters, and then the classifier is designed based on the cluster centers and membership degree function. Experimental results show that the proposed method has high recognition accuracy, compared with some common classification algorithms. Moreover, the proposed method recognizes lower limb activities based on small samples, which not only shortens the time for the classifiers to perform tasks, but also reduces the storage requirements of the system. 2.Existing research on activity recognition mainly analyzes some simple basic activities, and lacks discussion on the activities which involve a combination of upper limb and lower limb movements. A hierarchical method based on neural networks is proposed to divide complex concurrent activity recognition into two stages. At the lower level, one neural network is used to classify different lower limb activities. At the upper level, the upper limb movements are modeled by using multiple neural networks, and then the specific concurrent activity is inferred. The proposed hierarchical method divides the overall classification problem into two sub-problems, which reduces the complexity of the decision boundary. The performance is superior to some single-level methods. 3.In complex activity recognition, it is necessary to fully explore the correlation between activity time series data. To this end, a statistical method based on parallel hidden Markov model (PHMM) is proposed in this study, describing the continuation and transition of activity through a set of feature vector sequence. The time-domain and frequency-domain features are extracted as observation vectors, and the principle component analysis (PCA) method is used to reduce the dimension. To classify specific activities, PHMM makes the final decision by combining the probabilities of two channels. Experimental results show that, the proposed method recognizes activities with a higher accuracy, compared with some single-frame classifiers. 4.Existing research on sensor-based activity recognition mainly analyzes single-user activities, and lacks discussion on two-body interactive activities. In this study, some common interactive activities in people’s daily life are investigated, and the dynamic time warping (DTW) and Markov logic network (MLN) are employed to recognize them. Specifically, the DTW method is used to identifiy the actions of a single person, and the MLN is used to model interactive activities through semantics. The MLN assigns an associated weight to each rule, such that it has a certain ability to correct errors at making decisions. This effectively improves the recognition accuracy of two-body interactive activity. Key Words: Wearable Sensors; Human Activity Recognition; Recognition Accuracy; Feature Extraction; Principal Component Analysis

作者相关
主题相关
看过该书的人还在看哪些书