Questioning the Anisotropy of Pedestrian Dynamics: An Empirical Analysis with Artificial Neural Networks
Round 1
Reviewer 1 Report
Well presented.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The Authors propose to use neural networks to predict pedestrians' speed depending on the assumed configurations of pedestrians in relation to predecessors and followers. The Authors provide experiments proving advantages of their approach.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper is well organized and easy to follow. I agree to accept the current version.
Author Response
Comment 1 This paper is well organized and easy to follow. I agree to accept the current version.
Response Many thanks for the positive comment. Regarding improvement of the manuscript, we explained the
concept "anisotropy" in pedestrian interaction behaviour throughout the article (abstract, introduction,
related work, results and analysis, and conclusion). Also, we provided a new empirical analysis results
using homogeneous gender groups (females, and males alone) to enrich the paper’s contribution (see
section "Results and Analysis").
Author Response File: Author Response.pdf
Reviewer 4 Report
The paper is well-written with a clear idea of the problems. It presents nicely the background, methods, and results & discussion. However, the authors are suggested to modify the Conclusion and the quality of English.
Author Response
Comment 1 The paper is well-written with a clear idea of the problems. It presents nicely the background, methods,
and results & discussion. However, the authors are suggested to modify the Conclusion and the quality
of English.
Response Many thanks for the positive comment. We have revised the quality of English in the conclusion.
Besides, we highlighted the contribution of our research, and further explain the empirical analysis
results.
Author Response File: Author Response.pdf
Reviewer 5 Report
First, I would like to thank the authors for the paper they wanted to submit for review.
This is a well-written paper that addresses the problem of the dynamics of the movements of a single-line flow of pedestrians by analyzing situations that are referred to as high density (distances between pedestrians less than one meter). The authors examine the possibility of predicting the velocity of a given pedestrian by considering different parameters that define the distances between a subject pedestrian and other components walking in the front and behind.
The authors deal with an interesting topic in the general aspect of the study of crowd dynamics and in the particular one of analyzing the mutual influences between positions in the flow, both forward and backward. From this point of view, the literature review proposed by the authors is detailed and updated for the topic.
From reading the paper, the purpose of the work, the research question, and the message the authors intend to convey to the reader appear clear. However, here are some observations that I hope will help the authors clarify and improve their work.
Obs. 1
A first doubt concerns the meaning of anisotropy, which the authors call into question already with the title. They considered the longitudinal travel direction, examining what happened both in front of and behind each row component. Since there is a single row, transversal phenomena are not considered and therefore the concept of anisotropy should be better detailed and highlighted in line with the actual outlines of the study.
Obs. 2
Another doubt also concerns the adequacy of the title. In addition to citing anisotropy (with the necessary insights I tried to indicate above), the paper's title indicates the theme as "questioning the anisotropy". Considering that the authors do not question anisotropy, perhaps I think it would be more appropriate to make it clear from the title that their purpose is to investigate the opportunity to examine not only the direction (in this case, only the longitudinal one) of the interactions concerning the trajectory but also the position (i.e. concerning pedestrians walking in front but also to those walking behind).
Obs. 3
Just an observation is for the experimental data used for the different neural networks. The authors indicate data from an experiment in Palestine. Even if a reference is given in the paper, the 12th on the list, it would be better to give the reader some information about the experiment and how it was carried out, highlighting its main characteristics. Some of these aspects, such as the pedestrians' gender composition, are briefly called into question but not adequately clarified. In my opinion, this is important for the clarity of the work done. In particular, what is briefly written on lines 128 - 134 should be better detailed.
Obs. 4
Again relating to the experimental data, the authors refer to the use of data relating to a mixed flow of men and women. This aspect should be further clarified and specified. In this regard (lines 164 and 165), the authors wrote that, in the experiments, pedestrians were distributed in an orderly manner. It seems to be in such a way as to alternate men and women (single row of -man-woman-man-woman ---). If so, without prejudice to the need to better detail the circumstances and characteristics of the experiment, this alternation could have effects on the assessments conducted. I ask the authors if they have analyzed this aspect, whether differences emerge if we consider that the subject pedestrian can be a man or a woman, or other considerations that can emerge considering gender. These considerations could greatly enrich the paper's contribution or in any case, be considered as limits of this state of progress and proposals for further future investigations.
Obs 5
Another observation concerns the mean squared error (MSE) used to compare actual and simulated speeds in the training and test datasets. Figure 4 shows the box plot with the relative MSE calculated with D algorithms as a reference case. It is necessary to clarify how the values that produce the MSE distributions summarized in the figure are determined, based on which the observations reported by the authors are proposed.
Obs. 6
Finally, just a remark about conclusions. I advise the authors to increase the effect of the conclusions, enhancing the results obtained in comparison with what is usually analyzed and highlighted by the literature. As indicated in other observations, it would still be necessary to highlight the analysis limits based on which other research paths can be set. A suggestion on the actual uses of the results achieved is always an excellent way to consolidate the reader's interest and enhance the contribution that work provides to real-life applications.
Author Response
Comment 1 First, I would like to thank the authors for the paper they wanted to submit for review. This is a well-written paper that addresses the problem of the dynamics of the movements of a single-line flw of pedestrians by analyzing situations that are referred to as high density (distances between pedestrians less than one meter). The authors examine the possibility of predicting the velocity of a given pedestrian by considering different parameters that defie the distances between a subject pedestrian and other components walking in the front and behind. The authors deal with an interesting topic in the general aspect of the study of crowd dynamics and in the particular one of analyzing the mutual inflences between positions in the flw, both forward and backward. From this point of view, the literature review proposed by the authors is detailed and updated for the topic. From reading the paper, the purpose of the work, the research question, and the message the authors intend to convey to the reader appear clear. However, here are some observations that I hope will help the authors clarify and improve their work.
Response We would like fist to acknowledge the referee for the effort in studying our article and the rich, useful, and constructive remarks. Please fid below a detailed response to each of the comments and corresponding modifications we have done in the manuscript.
Comment 2 A fist doubt concerns the meaning of anisotropy, which the authors call into question already with the title. They considered the longitudinal travel direction, examining what happened both in front of and behind each row component. Since there is a single row, transversal phenomena are not considered and therefore the concept of anisotropy should be better detailed and highlighted in line with the actual outlines of the study.
Response Thanks for highlighting this point. Indeed, the concept of anisotropy in pedestrian interaction behavior is central to our analysis. We refer to anisotropic interaction behavior in the models in which the dynamics solely depend on the situation in front, i.e., in the direction of motion. Most pedestrian dynamic models include a vision-field and other mechanisms based on the bearing angle, making the interaction highly anisotropic, i.e., depending on the motion direction. In contrast, an isotropic model would assume that all pedestrians around, in front, on the sides, or behind, may influence the dynamic. We agree that our interpretation of anisotropy was not sufficiently detailed and that the manuscript, in its present form, may be confusing for the readers. To better explain how we understand the pedestrian behavior anisotropy in our study, we have in the revision of the manuscript introduced and explained the notion of anisotropic effects in pedestrian behaviors in the abstract, the introduction Sec. 1 (end of the third paragraph), the related work Sec. 2 (end of page 3), and Sec. 5 (highlighted).
Comment 3 Another doubt also concerns the adequacy of the title. In addition to citing anisotropy (with the necessary insights I tried to indicate above), the paper’s title indicates the theme as "questioning the anisotropy". Considering that the authors do not question anisotropy, perhaps I think it would be more appropriate to make it clear from the title that their purpose is to investigate the opportunity to examine not only the direction (in this case, only the longitudinal one) of the interactions concerning the trajectory but also the position (i.e. concerning pedestrians walking in front but also to those walking behind).
Response As mentioned in the response to the previous comment, the concept of anisotropy in pedestrian interaction behavior is central to our analysis. This is a reason why the term anisotropy is mentioned in the title. Our results focus on minimal single-fie scenarios for which isotropic interaction behavior directly depend on the use of the distances in front and behind for the interaction. Predictions based solely on the distances in front (that is classical for pedestrian models) would be considered anisotropic. While the behavior is supposed isotropic if the prediction relies on both distances ahead and behind. The results clearly show that isotropic interaction models allow significant prediction improvement. We agree that our use of anisotropy may be misunderstood and require to be better explained. Therefore, we have provided an important effort to better explain what we mean by the anisotropy notion in the abstract, the introduction, and later in the manuscript. Please read the response to the previous comment for details.
Comment 4 Just an observation is for the experimental data used for the different neural networks. The authors indicate data from an experiment in Palestine. Even if a reference is given in the paper, the 12th on the list, it would be better to give the reader some information about the experiment and how it was
carried out, highlighting its main characteristics. Some of these aspects, such as the pedestrians’ gender composition, are briefl called into question but not adequately clarifid. In my opinion, this is important for the clarity of the work done. In particular, what is briefl written on lines 128 - 134 should be better detailed.
Response Thank you for your comments. We further detailed the description and added a figure for the experiments. Please read the last revised part in Sec. 3.
Comment 5 Again relating to the experimental data, the authors refer to the use of data relating to a mixed flw of men and women. This aspect should be further clarifid and specifid. In this regard (lines 164 and 165), the authors wrote that, in the experiments, pedestrians were distributed in an orderly manner. It seems to be in such a way as to alternate men and women (single row of -man-woman-man-woman —). If so, without prejudice to the need to better detail the circumstances and characteristics of the experiment, this alternation could have effects on the assessments conducted. I ask the authors if they have analyzed this aspect, whether differences emerge if we consider that the subject pedestrian can be a man or a woman, or other considerations that can emerge considering gender. These considerations could greatly enrich the paper’s contribution or in any case, be considered as limits of this state of progress and proposals for further future investigations.
Response Thank you for you suggestions. We analyzed a new dataset for homogeneous gender groups (male, and female alone)(see section 3, and 5). We conclude that isotropic interaction behavior is also pronounced in females’ and males’ experiments, especially in the female group (N=20). However, because of the small dataset available to conduct this analysis, further empirical analysis with more data samples is required to emphasize the influence of distance behind on the pedestrian’s speed for homogeneous gender groups.
Comment 6 Another observation concerns the mean squared error (MSE) used to compare actual and simulated speeds in the training and test datasets. Figure 4 shows the box plot with the relative MSE calculated with D algorithms as a reference case. It is necessary to clarify how the values that produce the MSE distributions summarized in the figure are determined, based on which the observations reported by the authors are proposed.
Response In the analysis, we systematically perform a cross-validation to train the algorithms and test the algorithm. Such approach allow taking into for account for under- and overfitting features and to determine
the optimal algorithm complexity (i.e., optimal number of layers and neurons of the networks). To furthermore evaluate whether the prediction errors are precise and eventually determine whether differences between two errors are statistically significant, we repeat the cross-validation in a bootstrap loops by subsmapling randomly the training and testing sets. This allow us to obtain a distribution of the errors instead of a punctual estimate. We summarise the error distributions using boxplot diagrams in the manuscript. We have now provided in the fist paragraph of Sec. 5 more details about our
methodology as well as further references.
Comment 7 Finally, just a remark about conclusions. I advise the authors to increase the effect of the conclusions, enhancing the results obtained in comparison with what is usually analyzed and highlighted by the literature. As indicated in other observations, it would still be necessary to highlight the analysis limits based on which other research paths can be set. A suggestion on the actual uses of the results achieved is always an excellent way to consolidate the reader’s interest and enhance the contribution that work provides to real-life applications.
Response Many thanks for the comment. We have revised the conclusion and highlighted the contribution of our research, and further explain the empirical analysis results we have.
Author Response File: Author Response.pdf
Round 2
Reviewer 5 Report
I thank the authors for the new version of the paper and for providing all the explanations to my questions by extending. I consider this work by the authors as a well-done, well-written and interesting paper.
