Research on Pedestrian Avoidance Behavior for School Section Based on Improved BP Neural Network and XGboost Algorithm
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe aim of this work is to use 3 machine learning algorithms (BP neural networks, GA-BP neural networks and XGboost) to predict pedestrian avoidance behavior near school zones and to identify the most influential factors. The accuracy and precision of the algorithms are compared, and the best one is identified (GA-BP neural network).
Dear authors, please consider my suggestions and comments:
Abstract
- Please remove the sentence 'In 2024, a major traffic accident 11 occurred at the entrance of an elementary school in Shandong, where a vehicle collided 12 with pedestrians, drawing nationwide attention and posing a serious threat to people's 13 property and lives.' from the abstract, as it is not the appropriate place to report this type of information.
- The authors must clearly state the gap/innovation of the approach presented and expand a little more the main results (risk factor and how can be used in road safety decision).
Keyworks
- It's strange that the authors don't mention pedestrians, road accidents or schools in the keywords. These should be included.
Introduction
- The literature review should be much improved.
- Curiously, the presentation of the gap addressed at the end of the introduction section is accompanied by quotations. Revise to eliminate these quotations.
Data collection
- The description of the data collected should be improved, e.g. presenting the total time of the images collected in hh:mm; mentioning at the beginning that the data were collected from 6 schools; including a table with all the variables, their coding and the description of the coding and units; including the number of cases in the database (sample size used), etc.
- Are the speeds shown in line 111 of page 3 pedestrian or vehicle speeds? passenger cars or electric bicycles? in which units was it measured? In the Excel file of the database, all SV and OV values are less than or equal to 1. Have they been normalised? What about the STV, OTV and W values? Include this information in the text, as some of the explanatory variables are binary and others are continuous.
- Figure 1, page 4: The 6 school sites should be described, i.e. whether they have pedestrian crossings, pavements, central medians, etc.
- On what basis were the explanatory variables selected? This aspect is important and the rationale for the 12 variables considered should be justified.
Materials and Methods
- The organisation of the information presented in this section should be improved.
- The descriptions of the algorithms are very similar to those presented in other documents, so they should be greatly improved. The parameters of all equations presented should also be included. Several equations are presented without them, for example equations (1) to (4).
- In Figure 4, it is not clear how the 'Best weights' and 'Population' links between BP-NN and GA are made. Improve.
- Figure 5 on page 9 is missing. Amend.
Comparison analysis of prediction models
- It is not necessary to repeat the information presented in Tables 2 and 3 in the text, just refer to these tables (lines 295-299).
- In the graphs in Figure 6, it is preferable to use the same vertical scale to make it easier to compare and interpret the results.
- Why weren't the ROC values and feature importance results presented for the BP and GA-BP algorithms? These values should be included to allow comparison with the XGBoost results.
Conclusions
The conclusions should comment on the results obtained, in particular the identification with XGBoost of the most important variables, namely STV, STO, SV, SO, since it is expected that traffic volume and speed will influence pedestrian behavior. Finally, what is new about these results and how can they be used to support decision making?
General aspects
- All acronyms must be accompanied by their full meaning the first time they appear. Throughout the text there are several acronyms without their meaning. Proofread the entire manuscript.
- All figures and graphs should be enlarged so that they can be read properly. The titles of the vertical and horizontal axes and the units must be included on all graphs.
- Increase of the number of bibliographical references.
The English could be improved to more clearly express the research.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsThis manuscript presents a real-time video study to reduce the likelihood of road accidents around a primary and secondary school in the Pudong New Area, Shanghai, via a neural network approach.
The manuscript is interesting, although the following elements of weakness:
i) literature review is poor: the problem has been fully addressed in literature according several points of view. For example, start considering early approaches in: Appolloni, L.; Giretti, A., Corazza, M.V.; D’Alessandro, D. Walkable Urban Environments: An Ergonomic Approach of Evaluation. Sustainability 2020, 12(20), 8347. doi: 10.3390/su1112346 (neural network approach); Al-Najjar, E. et al. Addressing safety issues along the way to school: Qualitative findings from Jerash camp, Jordan, Journal of Transport & Health, 26, 2022, 101370, doi: 10.1016/j.jth.2022.101370 (specific built environments); Appleyard, B. Livable streets for schoolchildren: a human-centred understanding of the cognitive benefits of Safe Routes to School. Journal of Urban Design, 1-25. 2022. doi: 10.1080/13574809.2022.2070145 (walkability issues), as a start. Please, include Chinese and non-Chinese sources.
ii) What is the replicability/transferability of the proposed methodology? How can readership replicate the reported study elsewhere. Please, create a discussion section where "transferability" is elaborated.
iii) Figures are small, please enlarge them
iv) text from 324 to 338 is like a kind of bullet point list; to align with the rest of the text, turn it into a narrative. Ditto for 373 to 404
v) conclusions look like a proxy for a short summary. Highlight instead advances and limitations of the presented work, and next spteps
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
Comments and Suggestions for AuthorsThe article presents the following concerns:
- The quality of the figures needs to be improved.
- Please add hyperlinks to references, tables, and figures.
- Ensure all variables and parameters in the equations and figures are correctly referenced and described in the main text.
- Add a brief description between section and subsection titles.
- Add a header to the columns of Tables 2, 3, and 4.
- Several typos need to be fixed, including the one on line 196.
- Add a comparative table with the main results of this work compared to the state of the art.
- The exact number of pedestrian crossing events analyzed is not specified. The duration of the videos is mentioned (15,120 seconds), but the number of crossing events or avoidance instances is not identified. This makes it difficult to assess whether the sample is sufficient to train machine learning models without the risk of overfitting or insufficient data. Please specify.
- Traffic and vehicle speed variables will be considered, but other important contextual factors, such as lighting, weather conditions, traffic signals, or the presence of traffic enforcement officers, will be ignored. These can significantly affect pedestrian behavior. Please justify.
- Pedestrian avoidance is coded as one and non-avoidance as 0, but it is not detailed how pedestrian avoidance is determined. It is unclear whether the labeling was performed manually, using well-defined objective criteria, or whether automated tools were used to identify these behaviors. Justification is required.
- It is mentioned that the network architecture includes a hidden layer with 15 neurons, but no justification is given for choosing this number. No cross-validation experiments are presented to determine the best-fitting architecture.
- A genetic algorithm is used to optimize the neural network's initial weights, but it is not explained whether the optimization affects only the weights or also the network architecture. Furthermore, the choice of genetic algorithm parameters, such as population size (16), crossover probability (0.4), and mutation (0.5), is not justified.
- Although it is a robust model, it is not specified whether hyperparameter tuning was performed using cross-validation or if the values ​​used were arbitrary. There is no discussion of how the algorithm handles class imbalance, if any.
- The models' training time and computational efficiency, which are relevant for their implementation in real-world scenarios, are not mentioned.
- The F1 score is the main metric, but a detailed analysis of the confusion matrix regarding false positive and false negative costs is not reported. In a traffic safety problem, these errors can have different consequences (for example, wrongly predicting that a pedestrian will avoid a vehicle could be more serious than the opposite).
- The XGBoost feature importance analysis mentions traffic volume as the most relevant factor, but no causal analysis or robustness tests are provided. There could be an evaluation with another variable not considered.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have taken the suggestions and comments into account. They have significantly improved the manuscript.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors met the revision requirements. Manuscript fit for publication
Reviewer 3 Report
Comments and Suggestions for AuthorsThe manuscript can be published