Next Article in Journal
Driving Equity: Can Electric Vehicle Carsharing Improve Grocery Access in Underserved Communities? A Case Study of BlueLA
Previous Article in Journal
Mitigation of Voltage Magnitude Profiles Under High-Penetration-Level Fast-Charging Stations Using Optimal Capacitor Placement Integrated with Renewable Energy Resources in Unbalanced Distribution Networks
 
 
Article
Peer-Review Record

Comparative Analysis of Traffic Detection Using Deep Learning: A Case Study in Debrecen

Smart Cities 2025, 8(4), 103; https://doi.org/10.3390/smartcities8040103
by João Porto 1,*, Pedro Sampaio 1, Peter Szemes 2, Hemerson Pistori 1,† and Jozsef Menyhart 2
Reviewer 1: Anonymous
Reviewer 2: Anonymous
Reviewer 3: Anonymous
Reviewer 4:
Smart Cities 2025, 8(4), 103; https://doi.org/10.3390/smartcities8040103
Submission received: 22 April 2025 / Revised: 17 June 2025 / Accepted: 19 June 2025 / Published: 24 June 2025
(This article belongs to the Section Smart Transportation)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed a relevant and current topic in the applications and needs of the time, but in order to emphasize the individual scientific connection, below are presented some of recommendations and issues which need to be addressed from authors:

 

  1. The authors needed to expand dataset diversity variation in weather conditions: different times of day, Images during nighttime, rain, fog, or sun glare, seasonal variation (summer vs. winter traffic).
  2. Clarify data augmentation and preprocessing

          - Describe whether and how data augmentation was used?

           - Provide more detail on image resolution, annotation format, and preprocessing steps.

  1. Improve the quality of figures and visualizations
  2. Include Model Weights / Code Availability

-To improve reproducibility and impact, upload: trained model weights, code for preprocessing, training, and evaluation.

  1. Expand and improve the conclusion section

Clearly stated:

  • Which architecture is best for which task
  • How findings support future work in smart city planning
  • Limitations and proposed solutions (e.g., better data, more training time).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors
  • The abstract should clearly discuss in brief about the contributions of the article to the state of the art.
  • Discuss in brief about the main contributions of this work in the introduction section.
  • The authors should summarize the findings from the related works in teh form of a table or a paragraph, clearly highlighting the limitation of th state of the art in terms of traffic detection.
  • Did the authors considered imbalance in teh dataset? If so, how did they overcome this issue, if not, the authors should employ some techniques to address imbalanced dataset.
  • How are the hyper parameters chosen for the machine learning algorithms considered?
  • The authors should enhance the results analysis by analysing the computational complexity of the considered algorithms.
  • ALso, it is recommended that the authors perform a statistical testing to prove their hypothesis.
    • The authors can include the importance f Tiny ML/oneshot learning for small datasets and also explainable AI for justification or explaianbility of the results.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This is a comparative research paper with solid technology, clear structure and detailed experiments. It gives a new urban traffic dataset debstreet, and compares the performance of four mainstream architectures (yolov8, faster r-cnn, Detr, Sabl) in urban vehicle detection tasks. In this paper, the standardized evaluation index is used, and nine indexes are selected, which are map (average precision mean), map (average precision mean), map (average precision mean), map (average precision mean) and map (average precision mean) mAP@50 (IoU ≥ 50%)、 mAP@75 (IOU ≥ 75%), precision (precision), recall (recall), F1 score, MAE (mean absolute error), RMSE (root mean square error), Pearson correlation coefficient (R), and the performance of the model was systematically evaluated from the perspectives of generalization, mobility and adaptability.

From the second part of the specific experiment, its advantages are: the selection of data sets is reasonable (widely used ua-detrac and local data source debstreet); Model selection also covers the current mainstream models; In terms of index selection, it should be as perfect as possible, including classification and regression, which is helpful to evaluate its performance from two aspects of detection and counting; The experimental process adopts 10 fold cross validation and multi-stage migration test, which is systematic and rigorous. The number of images in the debstreet dataset is relatively small, only 682, but the number of labeled vehicles reaches 3256, which makes up for this to some extent. However, considering the large traffic flow at the intersection mentioned in the article, it is recommended to prepare more data sets. In the penultimate paragraph, the superparameters used in the training process (learning rate, number of training rounds, batch size and early stop strategy) are mentioned, which are respectively represented by symbols n, m, O and P, but their specific values are not indicated. They can be added for subsequent replication.

The third part conducts a comparative test from three stages, comprehensively evaluates four different models, tests the plasticity of the model, trains the model in the universal data set ua-drac and the localized data set debstreet (without fine tuning), and finally further improves the performance of the model through fine tuning. The three stages are reasonably designed and logically progressive; The analysis covers detection performance, counting ability, generalization ability and other dimensions; Several evaluation indexes and statistical test methods are used; The advantages and disadvantages of each model are analyzed in place, which is helpful for subsequent selection. In the second stage, it was mentioned that the accuracy of yoyov8 is better than that of fast r-cnn, but the number of missed detections is large (low recall). It is suggested to make a supplementary analysis of why this situation occurs in yoyov8, whether it is the reason of the yoyov8 model itself or the reason of parameter setting, resolution, etc. For the third stage, only the fine-tuning is mentioned, and the specific adjustment of the unit description can be considered for supplement, so that it can be reappeared later.

Since this paper focused on vehicle detection in urban environment, is there any contributions to urban scene understanding? Some experiments related to detection of vehicles in curved alleyways would help.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This paper entitled “Comparative Analysis of Traffic Detection Using Deep Learning: A Case Study in Debrecen” tackles the pressing challenge of accurate vehicle detection as a prerequisite for sustainable, data-driven urban mobility. The study offers a comprehensive review of state-of-the-art deep-learning architectures—Faster R-CNN, YOLOv8, DETR and SABL—and evaluates their performance on both a large public benchmark (UA-DETRAC) and a newly introduced, region-specific dataset (DebStreet) captured under varied weather conditions in Debrecen. By standardising nine evaluation metrics and employing a three-stage, cross-domain experimental design, the authors illuminate how attention mechanisms and transfer-learning strategies influence detection accuracy and vehicle-counting reliability. The findings underscore SABL’s superior mAP on in-domain tests and YOLOv8’s minimal counting error after fine-tuning, highlighting the trade-offs between precision and computational efficiency for real-time traffic monitoring. While the manuscript convincingly demonstrates the value of regional datasets and rigorous statistics, several aspects—such as dataset diversity, inference-time benchmarking and methodological novelty—could be further strengthened. Below are my specific remarks.

1.The entire paper only compares the existing detectors: YOLOv8, Faster R-CNN, DETR, and SABL. No new architectures or improvement strategies were proposed. It is suggested that the authors clarify their innovative technical points, combine attention mechanisms with lightweight design, or propose specific improvements for traffic scenarios to enhance the academic contribution of the paper.

  1. The dataset only contains 682 images, from a single intersection and a single "Vehicle" category. It is insufficient to cover scenarios with various climates, lighting conditions, and multiple target types. It is recommended to expand the sampling locations, time periods, and categories, and report the statistical data of the data distribution to enhance the generalization ability and external validity of the model.
  2. Three-stage - The settings of 10-fold cross-validation, ANOVA + Tukey test, etc. demonstrate high rigor. To enhance transparency, it is recommended to directly indicate the significant differences and p-values/effect sizes in the result table, and explain the details of weight reassignment on the specific impact on variance.
  3. The comparative experiment covers various mainstream detection architectures, cross-domain and fine-tuning scenarios, and has comprehensive indicator settings and sufficient result discussions. It is suggested to add practical engineering indicators such as inference time and resource consumption to make the comparison more applicable and valuable for reference, and to improve the overall performance evaluation of the model.
Comments on the Quality of English Language

 The English could be improved to more clearly express the research.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The recommendations are addressed.

Author Response

The authors would like to express their gratitude for the thorough review provided.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed all my queries. I have no further comments.

Author Response

The authors would like to express their gratitude for the thorough review provided.

Reviewer 3 Report

Comments and Suggestions for Authors

Most concerns were addressed. Some minors: since there was no experiments related to curved alleyways, some discussion with curved alleyways in street environment should be added or analyzed.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

I feel that you took the recommendations very serious and responded very good to these and where necessary, you adapted or elaborated the manuscript text. I've no further suggestions or comments.

Author Response

The authors would like to express their gratitude for the thorough review provided.

Back to TopTop