Dynamic Perception-Based Vehicle Trajectory Prediction Using a Memory-Enhanced Spatio-Temporal Graph Network
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsGeneral comments
l There are a lot of literature about using the recurrent networks or convolutional networks to predict vehicle trajectory. The related work needs a systematic review of relevant research.
l The issue summarized in the related work (“Although the above models consider interaction features within a certain structure, …the extracted environmental interaction features are not fully comprehensive”) should be more clearly described.
l The description of model construction is relatively brief, and the necessary formula derivation and concept description should be given.
l The evaluation indicators are relatively simple and the author should add different indexes, such as the specific degree of accuracy improvement.
l The selection of comparative models in the research is inconsistent with the content mentioned in the related work, and the selection criteria are also not explained. For example, the model is developed by combining Spatio-Temporal Graph Convolutional Network (STGCN) and attention-enhanced Long Short-Term Memory (LSTM) models, but this research lacks a comparison of prediction accuracy between the proposed model and the original model that is not combined.
l The discussion should be separated from the conclusion section.
Specifical comments
l L28, the expression "addresses time series issues" is inaccurate, and it lacks an explanation of time series issues.
l L192, the Experiments should be aligned to the left.
l L218, the article mentions conducting three ablation studies, while only two studies are found in the following text.
l L222-225, the specific definition of error evaluation indicator should be given.
l L228, Table 2, TrafficPredic and StarNet should be mentioned in the text.
l L233, there is no description of the indicator "Distance" and "circular lines" in Figure 3.
l The quality of some figures should be improved, e.g., Fig. 3, 4, 5, 6.
l L260-261, there is no obvious evidence to make a conclusion about "smaller model sizes and faster inference speeds".
Comments on the Quality of English LanguageEnglish expression needs to be modified.
Author Response
Dear Reviewer,
We have replied to your review comments in the attached Word document titled "Response to all reviewers." Please review at your convenience.
Best regards！
Author Response File: 
Author Response.docx
Reviewer 2 Report
Comments and Suggestions for Authorscomments attached
Comments for author File: 
Comments.pdf
Author Response
Dear Reviewer,
We have replied to your review comments in the attached Word document titled "Response to all reviewers." Please review at your convenience.
Best regards！
Author Response File: Author Response.docx
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper proposes a prediction method for vehicle trajectories, which is an integration of spatio-temporal graph convolutional network (GCN) and long short term memory (LSTM). Their method can treat not only individual trajectory, but also interactions affected from surrounding vehicles. Their proposed method is evaluated by using ApolloSpace dataset.
The reviewer felt this paper contains some problems.
[Problem1] The difference from the paper [28] is unclear.
The paper [28], "Grip: Graph-based interaction-aware trajectory prediction", has already proposed a similar approach, which integrates graph convolutional network and LSTM. Though the authors know Grip, this paper never mentions the difference between the proposed method and Grip.
The authors must explain what are the differences from the paper [28], and compare both methods in the experiment.
[Problem2] Explanation of assumptions (in Section 3) is not enough.
Some important assumptions in this paper are not explained.
* Who does predict vehicle trajectories? (each vehicle? cloud? road-side unit?) The execution environment assumed in this paper is not mentioned.
* Who does collect trajectories? Since a well-known dataset is used in their experiment, data collection part was omitted in this paper.
* The proposed method implicitly assumes that input data are vehicle trajectories on the same lane or lanes to the same direction. But, how to remove vehicle trajectories to the opposite direction (or other directions) is not mentioned.
[Problem3] Papers [26-43] in the references are not mentioned in the main body of the paper.
Papers [26-43] are included in the references, but not mentioned in the main body of the paper. The relationships between these papers and this paper are unclear. The authors must explain the relationships. Especially, "TrafficPredic" and "StarNet" in Table 2 need literature references.
Minor problems:
* Figure 3 and Figure 4 contain some Chinese texts.
* At Line 261, the authors wrote "smaller model size and faster inference speeds". Does this paper contain the experiment results corresponding to this sentence?
Author Response
Dear Reviewer,
We have replied to your review comments in the attached Word document titled "Response to all reviewers." Please review at your convenience.
Best regards!
Author Response File: Author Response.docx
Reviewer 4 Report
Comments and Suggestions for AuthorsDear authors of the article,
The material presented in the article is relevant and solves significant issues in intelligent transport systems (transport processes of megacities). The research presented by the authors and the proposed models have novelty and practical significance.
Today, due to the proliferation of autonomous transport systems, methods for predicting trajectories and movements are very relevant.
The authors pay special attention to the presentation of the graph structure spatial extraction and convolutional summing and the results of modeling and accuracy assessment.
The authors correctly presents the main problematic issues and reveals them. However, it is desirable to pay attention to the final presentation of results, to present in more detail. It is advisable to present a separate scientometric analysis of scientific articles.
The article is constructed logically correctly. The list of references is up-to-date. The article solves actual problem.
The presented analysis allows us to see at a higher and more detailed level transport model based on Spatio- Temporal Graph Convolutional Network (STGCN) with an attention-enhanced Long Short-Term Memory (LSTM) based Sequence to Sequence (Seq2Seq) encoder-decoder structure.
Author Response
Dear Reviewer,
We have replied to your review comments in the attached Word document titled "Response to all reviewers." Please review at your convenience.
Best regards!
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThe reviewer thinks the revised paper satisfactory fixed the problems which were pointed out by the reviewer at the last review.
