Detection Method for All Types of Traffic Conflicts in Work Zones
Round 1
Reviewer 1 Report
This study aims to build a detection method for all types of traffic accidents. One of the originalities the authors emphasized is detecting single-vehicle conflicts automatically. However, the authors did not represent how the proposed model is accurate and useful compared to the previous models. The output of this model is only the number of each conflict. Then accuracy is most important. I think the authors need major revision before acceptance.
My comments on what must be revised are as follows:
1. Please show the accuracy rate of SVM-based vehicle behavior identification model.
2. And show the usefulness of the proposed model compared to the previous models.
3. This whole model outputs the location and the number of the existing type of conflicts. Do the authors find new things through the model? Otherwise, this model is for practtical use, e.g. the advantage of automation and the decrease in the calculation time, etc. Please emphasize the originarity of this research except just detecting the single-vehicle conflicts.
Also, give minor-revision comments like the ones below(optional).
4. The type and the thresholds in this study are existing ones. I don’t think it is not good. But one of the advantages of the machine-learning/big-data-oriented model is to determine them. Why do the authors adopt the existing ones to the machine-learning model?
5. Regarding Table 8.(linear regression model), the number of evasive behaviors and traffic conflicts have spurious relationships because traffic volume is their common factor as Table6. and 7. show. If the authors think there are other factors, then please conduct other analysis such as path-analysis.
Author Response
This study aims to build a detection method for all types of traffic accidents. One of the originalities the authors emphasized is detecting single-vehicle conflicts automatically. However, the authors did not represent how the proposed model is accurate and useful compared to the previous models. The output of this model is only the number of each conflict. Then accuracy is most important. I think the authors need major revision before acceptance.
My comments on what must be revised are as follows:
- Please show the accuracy rate of SVM-based vehicle behavior identification model.
Reply: As is mentioned in Line 526, the best accuracy can be achieved at 90.1%.
- And show the usefulness of the proposed model compared to the previous models.
Reply: Thanks to the reviewer for the good suggestion. In the revised manuscript, we showed the usefulness of the proposed model, and verified its effectiveness.
Since we have mentioned in the introduction section that the previous models can only detect two-vehicle and multi-vehicle conflicts, and is not able to detect single-vehicle conflict, the usefulness and advantage of this method is to make up for this limitation and can be used to detect single-vehicle conflict. Meanwhile, this method can detect all types of conflicts at one time.
The effectiveness of using SSAM to detect two vehicle and multi vehicle conflict has been verified by many studies. Therefore, to verify the effectiveness of this method in all types of conflict detection, we only need to verify the accuracy of single vehicle conflict detection. The following method is adopted: Compare the number of single vehicle conflicts detected automatically with the number of single vehicle conflicts obtained from manual analysis of simulation video data. The results show that 19 of the 21 single vehicle conflicts are truly single vehicle conflicts through manual judgment, that is, the accuracy of the automatic single vehicle conflict detection algorithm is as high as 90%.
- This whole model outputs the location and the number of the existing type of conflicts. Do the authors find new things through the model? Otherwise, this model is for practtical use, e.g. the advantage of automation and the decrease in the calculation time, etc. Please emphasize the originarity of this research except just detecting the single-vehicle conflicts.
Reply: We emphasized in the introduction part of the revised manuscript. “The originality and contribution of this research have two aspects: (1)Reviewing the definition of traffic conflict based on evasive behavior, combining the two definitions, a method that can detect all types of traffic conflict is proposed, including single vehicle, two and multi-vehicle conflicts. (2) It realizes the full automation of conflict detection, and can efficiently complete the detection of all types of traffic conflicts based on the trajectory data obtained from microscopic traffic simulation and actual measurement, laying a good foundation for the fast safety assessment for work zones.”
Also, give minor-revision comments like the ones below(optional).
- The type and the thresholds in this study are existing ones. I don’t think it is not good. But one of the advantages of the machine-learning/big-data-oriented model is to determine them. Why do the authors adopt the existing ones to the machine-learning model?
Reply: Thanks for the good comment. Exactly, we also find the limitation of the threshold method for automatic segmentation in the verification of single-vehicle detection. As we have mentioned in Line 285, “in fact, some heuristic algorithms can realize the automatic setting of threshold, but since this is not the focus of this research, it won't be discussed further.” In further research, we will try the machine-learning/big-data-oriented model.
- Regarding Table 8.(linear regression model), the number of evasive behaviors and traffic conflicts have spurious relationships because traffic volume is their common factor as Table6. and 7. show. If the authors think there are other factors, then please conduct other analysis such as path-analysis.
Reply: We conducted the linear analysis to show the tight relationship between the evasive behavior and the SSAM output, which can be used to detect conflicts in an easy way in practice. As for other factors, we will conduct other analysis in further research.
Reviewer 2 Report
The intent of the research was to develop a methodology to identify all types of traffic conflicts in work zones. The study integrates the definitions of traffic conflict based on evasive behavior and the definitions of traffic conflict based on proximity and proposes an all-type of traffic conflict detection method applicable in work zones. The two and multi-vehicle conflict is analyzed by the classic SSAM method. The single-vehicle conflict detection method is based on previous research of the authors (Reference 2 of the paper) that determined that the definition of traffic conflict based on evasive behavior has the connotation of single-vehicle conflict and can effectively deal with detecting single-vehicle conflict. Evasive behavior is detected using vehicle micro-behavior data, automatic segmentation, SVM-based behavior identification, and threshold-based judgment methods with evasive behavior judgment based on acceleration.
The paper is written in good English. The structure of the paper is appropriate, the clarity is mostly at a sufficiently good level, and the authors presented the application of the proposed methodology in detail. With further improvement, regarding automatic segmentation and behavior identification, as the authors suggested in the paper, the proposed methodology has a potential to make a scientific and practical contribution to the field of traffic safety in work zones. I recommend the paper to be published after minor revision. Below you can find suggested corrections
1. Chapter 3.2 Proximity-based analysis: application of SSAM software and the procedure for Proximity-based analysis is poorly explained. For example, line 370 - Procedure starts without a proper introduction. Please provide a better explanation for the application of this method.
2. Please provide a link for reference 10, as it cannot be found on the Internet.
3. Figure 6: Put a description for the output of the SSAM software
4. The explanation for the abbreviation TRJ is not given in the text.
5. The font size in equations 1 and 2 appears to be different.
Author Response
- Chapter 3.2 Proximity-based analysis: application of SSAM software and the procedure for Proximity-based analysis is poorly explained. For example, line 370 - Procedure starts without a proper introduction. Please provide a better explanation for the application of this method.
Reply: This issue has been corrected and improved in the revised manuscript.
- Please provide a link for reference 10, as it cannot be found on the Internet.
Reply: This issue has been corrected and improved in the revised manuscript.
- Figure 6: Put a description for the output of the SSAM software
Reply: This issue has been corrected and improved in the revised manuscript.
- The explanation for the abbreviation TRJ is not given in the text.
Reply: This issue has been corrected and improved in the revised manuscript.
- The font size in equations 1 and 2 appears to be different.
Reply: This issue has been corrected and improved in the revised manuscript.
Reviewer 3 Report
Nice work and, indeed, you identified a gap. For readers not so deep in the subject it would be nice to be more plakative and brief. e.g. the formulas in 3.1.2. don´t really help to understand the outcome. They are import for you to calculate, but you need not to give all the details in the paper. Otherwise you should also explain why and how to apply the poission regression model with all output parameters and interpretations and statistics and assumptions behind it. In the abstract you should explain the abbreviations (line 15 and 17). In figure 2 you describe 11 basic behaviors and define them (above) e.g. with TL&C and so on. Later in 4.2. you are changing the order (line 502).
All the best.
Author Response
- Nice work and, indeed, you identified a gap. For readers not so deep in the subject it would be nice to be more plakative and brief. e.g. the formulas in 3.1.2. don´t really help to understand the outcome. They are import for you to calculate, but you need not to give all the details in the paper. Otherwise you should also explain why and how to apply the poission regression model with all output parameters and interpretations and statistics and assumptions behind it.
Reply: Thanks for the good suggestion. How to balance readers in different fields is really not easy to handle. In fact, in our earlier version, we didn’t explain these equations with such details. However, some colleagues expressed their expectation to see details in the subject.
- In the abstract you should explain the abbreviations (line 15 and 17).
Reply: This issue has been corrected and improved in the revised manuscript.
- In figure 2 you describe 11 basic behaviors and define them (above) e.g. with TL&C and so on. Later in 4.2. you are changing the order (line 502).
Reply: This issue has been corrected and improved in the revised manuscript.
Round 2
Reviewer 1 Report
Thank you for your revisions. I confirmed that the manuscript is well-revised. Now it is worth to be accepted.
