基于模型融合的人体行为识别方法研究

发布时间：2018-03-22 11:49

本文选题：行为识别　切入点：机器学习　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：人体行为识别是机器学习领域中一个重要的研究方向。随着传感器技术、移动互联网的快速发展和机器学习理论的成熟,基于移动设备传感器的人体行为识别技术越来越得到研究人员的关注。移动设备具有便携性、实时性、灵活性等优势,在人体行为识别领域得到广泛应用。人体行为识别技术具有研究价值和商业价值,具体应用场景有:人机交互、运动辅助、增强现实、智能监控等。人体行为识别是一个分类问题,因此,分类结果的准确率显得尤为重要。如果分类准确率不能达到应用可接受的范围,将会对用户产生负面的影响。分类准确率是由模型得到,一个模型的好坏对最终结果有很大影响。因此,本文首先通过实验分析,得到数据集非线性可分和难识别样本的问题会影响识别准确率的结论。然后,本文从模型方面进行优化,研究模型融合和模型参数选择来提高识别准确率。本文采集了移动智能设备下多种人体行为的加速度传感器数据,并对裸数据进行预处理和特征提取,得到可用于机器学习模型建模的数据样本。相较于DTW模板匹配算法,机器学习的分类模型具有更高的识别准确率。本文采用常用的机器学习模型对数据集进行建模分析。结果显示,常用的机器学习模型能够获得不错的识别准确率。尽管如此,常用的机器学习模型还有很大的提升空间。本文通过对实验结果分析发现,数据虽然经过预处理,但依然存在非线性可分和难识别样本的问题。单个模型往往很难处理这些问题。通过分析常用机器学习模型的优缺点,本文基于特征空间和模型预测两个角度进行优化,提出两种模型融合的方案。新模型相对于单个模型扬长避短,有效解决上述两个问题,提高识别准确率。最后,本文通过实验验证模型的正确性,并且针对模型的参数进行研究。
[Abstract]:Human behavior recognition is an important research direction in the field of machine learning. With the rapid development of sensor technology, mobile Internet and the maturity of machine learning theory, Human behavior recognition technology based on mobile device sensors has attracted more and more attention of researchers. Mobile devices have the advantages of portability, real-time, flexibility and so on. Human behavior recognition is widely used in the field of human behavior recognition. Human behavior recognition technology has research value and commercial value, the specific application scenes are: man-machine interaction, motion assistance, augmented reality, Intelligent monitoring and so on. Human behavior recognition is a classification problem, so the accuracy of classification results is particularly important. The accuracy of classification is obtained from the model, and the quality of a model has a great impact on the final result. It is concluded that the problem of nonlinear separable data sets and difficult to identify samples will affect the recognition accuracy. Then, this paper optimizes the model from the point of view of the model. Model fusion and model parameter selection are studied to improve the recognition accuracy. In this paper, the acceleration sensor data of various human behaviors under mobile intelligent devices are collected, and the naked data are preprocessed and feature extraction. Data samples can be used to model machine learning model. Compared with DTW template matching algorithm, The classification model of machine learning has higher recognition accuracy. In this paper, the commonly used machine learning model is used to model and analyze the data set. The results show that the commonly used machine learning model can achieve good recognition accuracy. There is still a lot of room for improvement in the commonly used machine learning models. Through the analysis of the experimental results, it is found that although the data is preprocessed, However, there are still problems of nonlinear separability and difficulty in identifying samples. These problems are often difficult to deal with by single model. By analyzing the advantages and disadvantages of commonly used machine learning models, this paper optimizes them from the perspectives of feature space and model prediction. Two models fusion scheme is proposed. Compared with a single model, the new model can effectively solve the above two problems and improve the recognition accuracy. Finally, the correctness of the model is verified by experiments. And the parameters of the model are studied.
【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;TP181

【参考文献】