Title Page
Abstract
Contents
Chapter 1. Introduction 11
1.1. Introduction 11
1.2. Aims and objectives of the thesis 13
1.3. Overreview of the thesis 13
1.4. Validation methodology and software/hardware tools 14
1.5. Contributions of the thesis 14
Chapter 2. Literature Review 15
2.1. Introduction 15
2.2. Handcrafted Representation-Based Approach 19
2.3. Learning-Based Action Representation Approach 19
2.3.1. Non-Deep Learning-Based Approaches 20
2.3.2. Deep Learning-Based Approach 21
2.4. Generative/Unsupervised Models 21
2.5. Discriminative/Supervised Models 22
Chapter 3. Human Action Recognition Using Transfer Learning 23
3.1. Introduction 23
3.2. Related Work 23
3.3. Methodology 24
Chapter 4. Human Action Recognition using Deep Belief Networks 25
4.1. Introduction 25
4.2. Related Work 25
4.3. Deep Belief Network 26
4.4. Structure Learning for Deep Networks 27
4.5. Architecture 28
4.5.1. Developing Architecture 28
4.5.2. Architecture Description 28
4.5.3. FC Layers 29
4.5.4. Architecture Methodology 29
Chapter 5. Datasets 30
5.1. KTH Human Action Dataset 30
5.2. IXMAS Dataset 30
5.4. UCF-101 Action Recognition Dataset 32
5.5. UCF Sports Action Dataset 32
Chapter 6. Implementation, Experiments and Results 36
6.1. Train Test Split 36
6.2. Preprocessing 36
6.3. Training 36
6.4. Testing 36
6.5. Computational Details 36
Results 37
6.7. UCF-101 37
6.8. KTH 37
6.9. UCF-Sports 37
Comparisons 38
6.10. UCF-101 38
6.11. KTH 38
6.12. UCF Sports 38
6.13. Action Quality Assessment 38
Conclusions 39
6.14. Performance 39
6.15. Future Works 39
References 40
Table 4.5.1. VGG Architecture 29
Table 5.1. [제목없음] 32
Table 6.1. UCF-101 Comparison 38
Figure 1.1. Diagram of a specific action recognition system. 12
Figure 2.1. Categorization by different level of Activities 17
Figure 2.2. Example of a kicking action using handcrafted representation-based approach 17
Figure 2.3. Example of a kicking action using learning-based approach 18
Figure 2.4. Traditional action representation and Recognition Approach. 19
Figure 2.5. Different layers of Convolutional Neural Networks 22
Figure 4.1. Diagram of DBN with visible layer V and four hidden layers. The blue arrows indicate the generative model, while red arrows indicate the direction of recognition. 26
Figure 4.3. The discriminative DBN model for classification with a visible layer V 27
Figure 4.5.1. VGG3D 28
Figure 5.1. One frame example of each action in Weizmann dataset 30
Figure 5.2. One frame example of each action from four different scenarios in the 31
Figure 5.3. One frame example for each action from five different camera views in IXMAS dataset 31
Figure 5.4. Exemplar frames for actions 1 to 57 from UCF-101 dataset 33
Figure 5.5. Exemplar frames for actions 58 to 101 from UCF-101 dataset 34
Figure 6.1. UCF-101 Train 37
Figure 6.2. UCF-101 Test 37
Figure 6.3. KTH Train 37
Figure 6.4. KTH Test 37
Figure 6.5. UCF-Sports Train 37
Figure 6.6. UCF-Sports Test 37