Human Action Recognition Using 3D-Convolution Neural Network

发布时间：2021-03-13 04:18

　　智能科技对现实环境中人类活动的敏锐分析为研究人员提供了广泛的应用领域,如对监控系统、客户理解、购物态度、正常或异常行为的分析等。然而,由于各种各样的局限性,如杂乱的背景、闭塞、视点变化等,要找到对行动的准确识别是一项具有挑战性的任务。我们必须牢记这些在视频中自动识别人类行为的局限性。实时自动识别HAR和非受控视频信息,如“监控视频”便是我们的主要关注点。近年来,研究人员试图提高基于视频的识别系统的准确度和精度,但并没有真正考虑到系统的效率。本研究主要考虑的是一个具有高精度值的髙效系统。另外,本文还重点研究了实时环境下的识别工具。此外,在复杂的环境中识别和分析人类行为更具有必要性与重要性。本研究的目的也在于区分正常行为与异常行为,并以系统的方式加以分类。综合研究表明,最近实现的分类是基于复杂度以及手工提取的原始输入特征。卷积神经网络具有直接作用于原始输入的能力,但也有处理二维输入的局限性。因此,本研究介绍了一种用于人体动作识别的新型三维卷积神经网络。此外,该方法是一种全自动的人类行为识别的深度模型。该学习过程并没有对人类行为进行分类的先验知识。因此本文建议方法包含两个步骤:第一步,应用三... 

【文章来源】：华中师范大学湖北省 211工程院校 教育部直属院校

【文章页数】：86 页

【学位级别】：硕士

【文章目录】：
Acknowledgements
Abstract
Chapter 1 Introduction
    1.1 General Background
    1.2 Problem Statement
    1.3 Significance of the Problem
    1.4 Contributions
    1.5 Objectives of the Research
    1.6 Thesis Outline
Chapter 2 Literature Review
    2.1 Designed Descriptor Based Methods
        2.1.1 Representation
        2.1.2 Classification
    2.2 Action Recognition Using Deep Models
        2.2.1 Convolution Neural Network
        2.2.2 Recurrent Neural Network
        2.2.3 Long Short-Term Memory Network
        2.2.4 3D-Convolution Neural Network
    2.3 Datasets for Human Action Recognition
        2.3.1 Simple Actions Datasets
        2.3.2 Complex Action Datasets
    2.4 Comparison of Our Approach with Related Work
        2.4.1 Cost and Efficiency
        2.4.2 Accuracy and Precision
Chapter 3 Proposed Method
    3.1 Representation and Classification of HAR
        3.1.1 Bag of Features Approach
        3.1.2 Fv Encoding Approach
    3.2 Theory of Convolution Neural Network
        3.2.1 Forward Propagation in Convolution Neural Network
        3.2.2 Backpropagation in Convolutional Neural Networks
        3.2.3 3D-Convolutional Neural Networks
    3.3 Proposed Method
        3.3.1 Step-1 Neural Network
        3.3.2 Step-2 Neural Network
Chapter 4 Experimental Results and Evaluation
    4.1 Feature Representation and Classification
    4.2 Brief Description of KTH and UCF11 Datasets
    4.3 Experiments on KTH and UCF11 Datasets
    4.4 Evaluation Protocol
    4.5 Results and Comparison
        4.5.1 Action Recognition on KTH dataset
        4.5.2 Action Recognition on UCF11 Dataset
    4.6 Advantages and Disadvantages of Using 3D-CNN
        4.6.1 Advantages of 3D-CNN
        4.6.2 Disadvantages of 3D-CNN
        4.6.3 Advantages of using RNN as Classifier
Chapter 5 Summary and Conclusions
    5.1 Summary
    5.2 Conclusions and Discussions
    5.3 Future Work
References
Appendix A Abstract and Summary
    A.1 Abstract
    A.2 Accepted Papers
    A.3 Environment Setting
        A.3.1 Windows Environment Setting
        A.3.2 Linux Environment Setting



本文编号：3079559