基于层次聚类特征选择和HF-SVM的活动识别技术

发布时间：2017-12-30 19:34

本文关键词：基于层次聚类特征选择和HF-SVM的活动识别技术　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：基于人体活动的智能计算是人工智能领域的一个重要研究方向,其目的是通过获取用户的状态和环境数据信息,为用户提供智能化应用服务。随着移动智能设备(如智能可穿戴设备)及其相关传感器等技术的飞速发展,基于移动智能设备的活动识别技术成为了研究的热点。由于移动智能终端在计算能力、存储空间和能量等硬件资源方面的限制,而传统机器学习模型需要巨大运算能力,基于移动智能设备的活动识别技术无法广泛应用。针对以上问题,本文在分析传统特征选择和支持向量机的基础上,提出了一种基于层次聚类的特征选择和硬件友好型核函数的支持向量机的多类别分类方法。论文的主要算法改进和成果如下:(1)提出了改进的基于层次聚类算法的特征选择算法。基于层次聚类的特征选择算法使用的评价函数基于互信息和关联系数。这不能适用于活动识别领域的连续型数据。本文基于皮尔逊相关系数和共享最近邻这两种度量改进了评价函数。使用改进的基于层次聚类的特征选择算法,完成了特征提取,降低活动识别技术在模型训练过程中的复杂度。(2)提出了基于硬件友好型核函数的SVM算法。传统的SVM算法在模型训练和算法应用中需要大量的指数运算。本文基于高斯核函数和拉普拉斯核函数提出了硬件友好型的核函数,既保持高斯核函数抗噪声的优点,也具有较小的计算代价。(3)提出了基于组合分类器的活动识别技术。活动识别是多类别分类问题。传统的多类别分类算法存在分类误差大的问题。本文提出了基于OVO策略的SV算法,将每个分类器的输出结果作用于Sigmoid函数,然后根据每个二值分类器的投票得到最终输出。本文提出的方法避免单独使用Sigmoid函数取最大值时易受到噪声的干扰而预测出错,也避免了多个分类器直接投票由于分类器权重相同导致的错误。本文从以上三个方面对活动识别技术进行研究,展开的实验分析显示文中提出的活动识别技术具有较高准确率,具有一定的实用价值。
[Abstract]:Based on Computational Intelligence of human activities is an important research direction in the field of artificial intelligence, its purpose is to obtain the user's state and the environment data, provide intelligent services for users. With the application of smart mobile devices (such as smart wearable devices) and rapid development of sensor technology, identification technology of mobile activities intelligent equipment has become a hot research topic. Because based on mobile intelligent terminal in computing power, storage space and power hardware resource constraints, while the traditional machine learning model requires huge computational ability, activity recognition technology of mobile smart devices can not be widely used on. To solve the above problems, based on the analysis and selection of support vector machine the traditional features, proposes a classification based on multi class support vector machine feature selection and hardware friendly kernel hierarchical clustering of the Methods. The main algorithm and results of this paper are as follows: (1) proposed an improved feature selection algorithm based on hierarchical clustering algorithm. Based on the evaluation of feature selection algorithm using hierarchical clustering function based on mutual information and correlation coefficient. Continuous data cannot be applied to this activity recognition field. In this paper, and the two shared nearest neighbor for improving the measurement of the evaluation function based on Pearson correlation coefficient. Using the improved feature selection algorithm based on hierarchical clustering, the feature extraction and recognition technology in model training activities reduce the complexity in the process. (2) proposed the SVM algorithm hardware friendly kernel function. Based on the traditional SVM algorithm needs a large number of index operations in the model training and the application of the algorithm. The Gauss kernel function and Laplasse kernel function is proposed based on kernel function hardware friendly, while maintaining the Gauss kernel function of anti noise The advantages, also has a smaller computational cost. (3) proposed activity recognition technology based on combining classifiers. Activity recognition is a multi class classification problem. Multi class classification algorithms exist for the traditional classification error. This paper proposes a SV algorithm based on OVO strategy, the results of the output of each classifier in Sigmoid then according to the value function, each of the two classifiers vote to obtain the final output. This method uses the Sigmoid function to avoid separate noise susceptible to maximum value and prediction error, but also avoid the multiple classifiers vote because classifier weights of the same mistake. The activity recognition technology research from the above three aspects the experiment shows that, activity recognition technology is proposed in this paper has high accuracy, and has certain practical value.

【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】