基于智能手机传感器数据的人类行为识别研究

发布时间：2019-02-24 15:02

【摘要】：人类行为识别是模式识别领域的研究热点,目前有基于视频数据和基于智能手机传感器数据两种研究方向。随着时代的快速发展,智能手机的普及化以及嵌入在智能手机中的传感器的多样化,使得运用智能手机传感器数据进行人类行为识别研究更加有实际意义。本文主要针对人类行为识别中出现的分类结果精度不准确,实验特征数据量大,分类效果不佳等问题,在真实数据的基础上,提出了基于多阶层连续隐马尔科夫模型的人类行为识别和基于稀疏局部保持投影结合随机森林集成分类器(Sparse Locality Preserving Projections and Random Forest,SpLPP-RF)的人类行为识别两种创新性算法,有效的解决了目前行为识别研究中遇到的困难。本文主要研究成果如下所示:(1)传统的连续隐马尔科夫模型在进行行为识别时,最终的行为识别准确率相对较低。基于人类活动的层次特点与传感器数据的时序性、多元性与连续性,本文提出了三阶层连续隐马尔科夫模型(Three-Stage Continuous Hidden Markov Model,TSCHMM)的人类行为识别新算法。实验结果表明所提出的算法不仅可以明显判别出活动的误分类别,而且解决了识别率低的问题,尤其是提高了易混淆活动的分类准确率。(2)首次将稀疏局部保持投影算法应用于连续隐马尔科夫模型的人类行为识别中。稀疏局部保持投影(Sparse Locality Preserving Projections,SpLPP)优化保留了邻域结构的数据集,并且相比于局部保持投影算法,可以从传感器数据中提取出更多有代表性的活动行为特征变量。以SpLPP作为降维方法的实验结果表明新算法效果明显。(3)由于集成分类器一般情况下分类效果优于单一分类器,所以已有一些研究采用随机森林(RandomForest,RF)分类器应用于智能手机传感器数据的人类行为识别上。但他们的方法没有充分利用较前沿的降维技术。因此,本文提出了用SpLPP进行降维,有效的解决了人类行为识别研究中的特征数量多的问题,降低了实验的时间复杂度,行为识别的总体识别率得到了显著的提高。同时,也比较了 SpLPP-RF和TSCHMM这两种算法,阐述了两种算法在性能上的差异与适用情况。
[Abstract]:Human behavior recognition is a hot topic in the field of pattern recognition. There are two research directions: video data and smart phone sensor data. With the rapid development of the times, the popularization of smart phones and the diversity of sensors embedded in smart phones, it is more meaningful to study human behavior recognition using smart phone sensor data. Based on the real data, this paper mainly aims at the problems of inaccurate classification results, large experimental data volume and poor classification effect, which appear in human behavior recognition. Two innovative algorithms for human behavior recognition based on multilevel continuous hidden Markov model and sparse local preserving projection combined with stochastic forest ensemble classifier (Sparse Locality Preserving Projections and Random Forest,SpLPP-RF) are proposed. Effectively solve the current behavior recognition research encountered difficulties. The main results of this paper are as follows: (1) the accuracy of behavior recognition in the traditional continuous hidden Markov model is relatively low. Based on the hierarchical characteristics of human activities and the timing, plurality and continuity of sensor data, a new human behavior recognition algorithm based on three-level continuous Hidden Markov Model (Three-Stage Continuous Hidden Markov Model,TSCHMM) is proposed in this paper. The experimental results show that the proposed algorithm can not only clearly distinguish the wrong classification of activities, but also solve the problem of low recognition rate. In particular, the classification accuracy of confusing activities is improved. (2) the sparse local preserving projection algorithm is applied to human behavior recognition of continuous Hidden Markov models for the first time. Sparse local preserving projection (Sparse Locality Preserving Projections,SpLPP) optimizes the neighborhood structure of the data set, and can extract more representative activity characteristic variables from the sensor data than the local preserving projection algorithm. The experimental results using SpLPP as a dimensionality reduction method show that the new algorithm is effective. (3) because the ensemble classifier is better than a single classifier in general, some researches have adopted random forest (RandomForest,). RF) classifier is applied to human behavior recognition of smart phone sensor data. However, their methods do not make full use of the more advanced dimensionality reduction techniques. Therefore, this paper proposes to reduce the dimension by using SpLPP, which effectively solves the problem of the large number of features in the research of human behavior recognition, reduces the time complexity of the experiment, and improves the overall recognition rate of behavior recognition significantly. At the same time, two algorithms, SpLPP-RF and TSCHMM, are compared, and their performance and applicability are discussed.
【学位授予单位】：浙江师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.4;TP212.9

【参考文献】