視頻中的人體姿勢、動作識別是人類行為自動分析理解的基本任務(wù)。無論在運動還是靜止情況下,獲取人體信息都必需進(jìn)行人體姿勢、動作識別。隨著機器學(xué)習(xí)的快速發(fā)展和深度學(xué)習(xí)技術(shù)的進(jìn)步,尤其是用于特征提取、分類或回歸的端到端深度神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu),成為提高圖像和視頻中姿勢估計和動作識別性能的關(guān)鍵。在本論文中,我們提出了使用深度卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行姿態(tài)估計和動作識別的新技術(shù),這是一種專門為二維特征提取而設(shè)計的深度神經(jīng)網(wǎng)絡(luò)。由于深度卷積神經(jīng)網(wǎng)絡(luò)能夠自動學(xué)習(xí)訓(xùn)練數(shù)據(jù)中的低級和高級特征,基于深度卷積神經(jīng)網(wǎng)絡(luò)的方法優(yōu)于此前基于特征工程的方法。由于在圖像識別的關(guān)鍵是根據(jù)所需任務(wù)提取相關(guān)特征,因此在我們提出的技術(shù)中,重點是如何利用新的深度卷積神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)來改進(jìn)特征提取。我們從數(shù)據(jù)類型和問題性質(zhì)兩個不同方面解決問題。首先,我們將深度圖像中的三維姿態(tài)估計和彩色圖像中的二位姿態(tài)估計視為回歸問題,在使用深度卷積神經(jīng)模型進(jìn)行端到端學(xué)習(xí)的過程中,我們將輸入圖像直接映射到姿勢位置。其次,我們同時使用深度圖像和三維姿勢數(shù)據(jù)來構(gòu)建提供不同類型的運動特征的兩個描述符,然后設(shè)計了三個深度卷積神經(jīng)網(wǎng)絡(luò)通道用于特征提取和動作分類。最后,作為一項補充... 

【文章來源】：上海交通大學(xué)上海市 211工程院校 985工程院校 教育部直屬院校

【文章頁數(shù)】：140 頁

【學(xué)位級別】：博士

【文章目錄】：
摘要
Abstract
List of Abbreviations
Chapter 1 Introduction
    1.1 Background
    1.2 Thesis Objective
    1.3 Challenges
        1.3.1 RGB-D-Based3D Single Person Pose Estimation
        1.3.2 RGB-Based2D Single Person Pose Estimation
        1.3.3 RGB-D-Based and Posture-Based Action Recognition
        1.3.4 Posture-Based Motion Quantification
    1.4 Contributions
    1.5 Thesis Structure
Chapter 2 Related Work
    2.1 Human Pose Estimation
        2.1.1 RGB-D-Based3D Pose Estimation
        2.1.2 RGB-Based2D Single-Person Pose Estimation
    2.2 Human Action Recognition
        2.2.1 RGB-D-Based Action Recognition
        2.2.2 Skeleton-Based Action Recognition
    2.3 Wearable and Wireless Sensor-Based Pose Estimation and Action Recognition
    2.4 Human Motion Capture and Motion Comparison
Chapter 3 3D Human Pose Estimation From a Single Depth Image
    3.1 Overview
    3.2 Human Pose Estimation Method
        3.2.1 Data Normalisation
        3.2.2 Convolutional Neural Network Model
    3.3 Experimental Results
        3.3.1 Training and Testing
        3.3.2 Comparison and Discussion
Chapter 4 Single-Person2D Pose Estimation Using Hybrid Refinement-CorrectionHeatmaps
    4.1 Overview
    4.2 Hybrid Refinement-Correction Pose Estimation
        4.2.1 Pose Refinement
        4.2.2 Pose Correction
        4.2.3 Heatmaps Fusion
    4.3 Experimental Results
        4.3.1 Training and Testings Settings
        4.3.2 MPII Dataset
        4.3.3 FLIC Dataset
        4.3.4 Influence of the Correction Network（CNet）
        4.3.5 Computation Complexity
Chapter 5 Action-Fusion:Human Action Recognition Using Depth Images andBody Postures
    5.1 Overview
    5.2 Action Recognition Method
        5.2.1 Action Descriptors
        5.2.2 Convolutional Neural Network Model
        5.2.3 Score Fusion
    5.3 Experimental Results
        5.3.1 Datasets Results Evaluation
        5.3.2 Processing Computation Complexity
Chapter 6 Effective3D Joints-Based Human Motion Quantification and SimilarityEvaluation
    6.1 Overview
    6.2 Method
        6.2.1 Motion Quantification
        6.2.2 Motion Comparison
    6.3 Experimental Results
        6.3.1 UTD-MHAD Dataset
    6.4 User Movements Comparison Study Using Kinect
        6.4.1 Comparison with Existing Methods
Chapter 7 Conclusion
    7.1 Summary
        7.1.1 General Summary
        7.1.2 Detailed Summary
    7.2 Limitations and Possible Improvements
    7.3 Proposed Future Work
Bibliography
Acknowledgements
List of Publications



本文編號：3114873