Title Page
Contents
List of Abbreviations 9
요약 12
Abstract 14
Ⅰ. INTRODUCTION 16
1.1. Research Background 16
1.2. Problem Definition 17
1.3. Research Purposes and Contributions 19
1.4. Related Works 20
1.5. Structure of the Thesis 25
Ⅱ. Traffic Sign Detection and Recognition System 26
2.1. Traffic Sign Detection 26
2.2. Traffic Sign Classification 27
2.3. Traffic Sign Augmentation 29
Ⅲ. Methodology 30
3.1. German Traffic Sign Recognition Benchmark 30
3.1.1. Collection of Dataset 30
3.1.2. Data Organization 32
3.2. Chinese Traffic Sign Dataset (CTSD) 34
3.2.1. Collection of Dataset 34
3.2.2. Data Organization 34
3.3. Data Preprocessing 36
3.4. Recognition Method Outline 36
3.5. Experimental Environment 37
Ⅳ. Proposed Architecture 39
4.1. ROIs Extraction Module 39
4.2. Dataset and Augmentation 43
4.2.1. GTSDB & CTSD Dataset 43
4.2.2. Augmentation 43
4.2.3. Flipping Method 44
4.2.4. Rotation and Projection Method 44
4.3. Multi-Task CNN 45
4.4. Early Stopping 48
4.5. Post Processing 50
Ⅴ. Experimental Results 51
5.1. GTSDR Dataset 51
5.2. CTSD Dataset 60
Ⅵ. Conclusions and Future Works 67
6.1. Conclusions 67
6.2. Limitations 67
6.3. Future Works 68
References 69
Table 1. Summary of existing literature on TSR. 24
Table 2. Four categories of traffic signs examples. 33
Table 3. Hardware environment for experiment. 37
Table 4. Software environment for experiment. 38
Table 5. Comparison of loss and accuracy value of the GTSDR dataset. 58
Table 6. Comparison of loss and accuracy value of the CTSD dataset. 65
Figure 1. Some common issues encountered by TSDS: (a) Occlusion, (b) Snow, (c) Color Fading, and (d) Undesirable Light. 18
Figure 2. Basic flow of traffic sign detection and classification systems. 28
Figure 3. Some images from GTSRB Dataset. 31
Figure 4. Chinese Traffic Sign Dataset. 35
Figure 5. ROIs module. 40
Figure 6. The MaxPool operation. 46
Figure 7. Model Summary. 47
Figure 8. An illustration of early stopping approaches. 48
Figure 9. Comparison of Loss and Accuracy with Early Stopping Epochs (GTSDR). 53
Figure 10. Comparison of Loss and Accuracy with 25 Epochs (GTSDR). 55
Figure 11. Comparison of Loss and Accuracy with 15 Epochs (GTSDR). 56
Figure 12. Comparison of Loss and Accuracy with 10 Epochs (GTSDR). 57
Figure 13. Experimental Result (GTSRB). 58
Figure 14. Confusion matrix of the GTSRB dataset. 59
Figure 15. Comparison of Loss and Accuracy with 10 Epochs (CTSD). 61
Figure 16. Comparison of Loss and Accuracy with 15 Epochs (CTSD). 62
Figure 17. Comparison of Loss and Accuracy with 25 Epochs (CTSD). 63
Figure 18. Comparison of Loss and Accuracy with early stopping Epochs (CTSD). 64
Figure 19. Experimental Result (CTSD). 65
Figure 20. Confusion matrix of the CTSD dataset. 66