Title Page
Abstract
Contents
Chapter 1. Introduction 18
1.1. Background and Motivation 18
1.2. Rule-based approach VS Data-driven approach 20
1.3. The feature extraction process of the YOLO 27
1.4. Back propagation in CNN 32
Chapter 2. Materials and Methods 36
2.1. The YOLOv5_Ours Network 36
2.2. Data Preparing and Processing 58
Chapter 3. Experimental and Results 72
3.1. Experimental Setup and Flowchart 72
3.2. Key Indicators for the experiment 75
3.3. Experimental loss function 77
3.4. Experimental results 87
3.5. Comparison with previous Object detection models 108
3.6. Application on Various Datasets 110
Chapter 4. Conclusions 113
References 115
Abbreviations 126
국문 초록 129
Table 2.1. Accuracy comparison table of ReLU, SiLU, and ELU activation functions (in ResNet) 48
Table 2.2. Classification of the dataset based on weather conditions 71
Table 3.1. Comparison table of the original YOLOv5 and the YOLOv5_Ours models 88
Table 3.2. Key Indicators of the YOLOv5_Ours Model 102
Table 3.3. Key Indicators of the YOLOv5_Ours Model (When the number of classes is 5) 104
Table 3.4. Comparison table of the original YOLOv5 and the YOLOv5_Ours models (When the number of classes is 5) 104
Table 3.5. Comparison table of the original YOLOv5 and the YOLOv5_Ours models (only VisDrone) 106
Table 3.6. Comparison of performance between previous models and YOLOv5_Ours 108
Table 3.7. Comparison of Results Obtained from Applying to Different Datasets 111
Table 3.8. Comparison of frames per second by model 112
Figure 1.1. Method of the rule-based approach (finding rules) 20
Figure 1.2. Method of the rule-based approach (finding edge) 21
Figure 1.3. Method of the data-driven approach 22
Figure 1.4. Method of the data-driven approach in YOLO 23
Figure 1.5. Method of data-driven approach in YOLO (label conversion) 24
Figure 1.6. Method of the data-driven approach in YOLO(within Images) 25
Figure 1.7. Method of the data-driven approach in YOLO(label component) 26
Figure 1.8. Method of feature extraction in YOLO(grid division) 27
Figure 1.9. Method of feature extraction in YOLO (create of anchor boxes) 29
Figure 1.10. Method of feature extraction in YOLO (one cell component) 30
Figure 1.11. Sum of final parameters (per image) 31
Figure 1.12. Back propagation algorithm architecture 33
Figure 2.1. Structure of the previous YOLOv5 model 39
Figure 2.2. Structure of the current YOLOv5 model 40
Figure 2.3. Graph comparison of the SiLU, ELU and ReLU activation functions 44
Figure 2.4. Graph comparison of the SiLU, ELU and ReLU derivative functions 45
Figure 2.5. Accuracy comparison graph of the ReLU, SiLU, and ELU activate functions (in ResNet) 47
Figure 2.6. Accuracy comparison graph of the ReLU, SiLU, and ELU activate functions (in Classification) 49
Figure 2.7. Accuracy comparison graph of the ReLU, SiLU, and ELU activate functions (a typical CNN structure using MNIST) 50
Figure 2.8. Comparison of the Conv and Conv_Ours layer (in the format suggested by the right figure) 52
Figure 2.9. Comparison of the original C3 and C3_Ours layer (in the format suggested by the right figure) 53
Figure 2.10. Comparison of the original SPPF and SPPF_Ours layer (in the format suggested by the figure below) 54
Figure 2.11. Flowchart of the first ConvELU layer (in the format suggested by the figure) 56
Figure 2.12. Structure of YOLOv5_Ours (in the format suggested by the figure) 57
Figure 2.13. Sample image taken on a clear day 60
Figure 2.14. Sample image taken on a cloudy day 61
Figure 2.15. Sample image taken on a rainy day 62
Figure 2.16. Sample image taken on a snowy day 63
Figure 2.17. Sample image taken on a evening 64
Figure 2.18. Sample image taken on a night 65
Figure 2.19. Sample image taken on a low altitude 66
Figure 2.20. Sample image taken on a high altitude 67
Figure 2.21. The process of converting drone-collected data samples to the YOLO format on a snowy day 69
Figure 2.22. The process of converting drone-collected data samples to the YOLO format on a Evening 70
Figure 3.1. The structure of the experimental method 74
Figure 3.2. Intersection over Union (IoU) definition 78
Figure 3.3. The value of Intersection over Union (IoU) in each situation 79
Figure 3.4. The value of GIoU in each situation 81
Figure 3.5. The process by which GIoU identifies an object 82
Figure 3.6. Normalized distance between bounding box and ground truth 84
Figure 3.7. The process by which the CIoU finds an object 86
Figure 3.8. Comparison graph of results for original YOLOv5 and YOLOv5_Ours model 88
Figure 3.9. Graph of result values for YOLOv5_Ours model (changes in key indicators according to the epochs of training) 90
Figure 3.10. Graph showing the precision-recall curve for the YOLOv5_Ours model 92
Figure 3.11. Detection Results of YOLOv5_Ours Model on a Clear Day 94
Figure 3.12. Detection Results of YOLOv5_Ours Model on a Cloudy Day 95
Figure 3.13. Detection Results of YOLOv5_Ours Model on a Rainy Day 96
Figure 3.14. Detection Results of the YOLOv5_Ours Model on a Snowy Day 97
Figure 3.15. Detection Results of the YOLOv5_Ours Model on at Evening 98
Figure 3.16. Detection Results of the YOLOv5_Ours Model on at Night 99
Figure 3.17. Detection Results of the YOLOv5_Ours Model at Low Altitude 100
Figure 3.18. Detection Results of the YOLOv5_Ours Model at High Altitude 101
Figure 3.19. Graph showing the precision-recall curve for the YOLOv5_Ours model (When the number of classes is 5) 105
Figure 3.20. Graph showing the precision-recall curve for the YOLOv5_Ours model (only VisDrone) 107
Figure 3.21. Comparison graph of result values for previous models and YOLOv5_Ours 109