Title Page
Abstract
Contents
I. Introduction 9
II. Related Studies 11
1. Mask R-CNN Overview 11
2. Vehicle Damage Volume Recognition 12
3. Traffic Accident Types Recognition 14
4. Problem Analysis 15
A. Vehicle Damage Volume Recognition 15
B. Traffic Accident Types Recognition 17
III. Proposed Algorithm 19
1. Design Philosophies 19
2. Algorithm Overview 20
3. Function of Vehicle Damage Volume Recognition 21
4. Function of Traffic Accident Types Recognition 22
IV. Experimentation and Performance Evaluation 24
1. Dataset Construction 24
A. Dataset Collection 24
B. Dataset Annotation 25
C. Dataset Training (Pre-trained Mask R-CNN model) 25
2. Experimental Environments 27
3. Evaluation Metrics 28
A. Confusion matrix 28
B. Accuracy 29
C. Precision 29
D. Recall 29
E. F1-Score 30
4. Experimentation Scenario 30
5. Performance Evaluation 31
A. Trained Dataset 31
B. Function of Vehicle Damage Volume Recognition 33
C. Function of Traffic Accident Types Recognition 34
C. Comparing the damage volume recognition result with previous Research 35
6. Discussion 37
V. Conclusions 39
References 40
Table 1. Experimental Environment Information. 27
Table 2. Parameter Value for Dataset Training. 28
Table 3. Summary of Vehicle Damage Volume Recognition. 37
Table 4. Summary of Traffic Accident Types Recognition. 38
Figure 1. Mask R-CNN Architecture. 12
Figure 2. Classification of Damage Types 12
Figure 3. Damage Types of Vehicles. 13
Figure 4. Three Common Types of Vehicle Accidents. 14
Figure 5. Automatic Vehicle Accident Detection. 15
Figure 6. Common Concept of Existing Researches. 16
Figure 7. Problem Analysis of Damage Types. 17
Figure 8. Automatic Traffic Accidents Detection. 18
Figure 9. Algorithm Design Overview. 20
Figure 10. Function of Vehicle Damage Volume Recognition. 21
Figure 11. Function of Traffic Accident Types Recognition. 22
Figure 12. Collection Process of Dataset. 24
Figure 13. Image Annotation Process. 25
Figure 14. Pre-trained Mask R-CNN Model. 26
Figure 15. Experimentation Scenario. 30
Figure 16. Performance Evaluation with Confusion Matrix. 32
Figure 17. Performance Evaluation of Trained Dataset. 32
Figure 18. Performance of Damage Volume Recognition. 33
Figure 19. Performance of Traffic Accident Types Recognition. 34
Figure 20. Comparing Performance with Previous Research. 35