Traffic Light Detection by Integrating Feature Fusion and Attention Mechanism

Round 1

Reviewer 1 Report

1. Abstract

1) The author should explain the existing problems of the existing methods of traffic light classification, and explain the research value of why NLP should be used to solve the corresponding problems.

2) The method in the abstract needs to explain the specific role of such network architecture and module design.

3) The discussion section of the experimental results should be specified about what kind of accuracy improvement has been achieved.

2. Introduction

1) When introducing the research value of traffic light classification, relevant references about traffic light classification should be cited to prove your point.

2) In addition to introducing the significance of traffic light classification research, it is also necessary to briefly introduce the main methods and development process of existing traffic light classification related research. And analyze the problems of these methods.

3) It is necessary to introduce the innovation points of this paper and what kind of problems can be solved.

3. Related works

1) The latest researches on the classification of traffic light signals needs to be introduced and analyzed.

4. Methodology

5. Experiments

1) Line347, the equation of AP and mAP should be illustrated.

2) The compared methods only includes Yolo. Other state of arts method for traffic light classification should be compared also. In addition, in table 3, why the precision,recall and f1score of SSD, Faster R-CNN and Yolo are missing. These methods should be compared completely.

3) In table 5, add some explanations for the bold values. Otherwise, the bold values are confused. The readers cannot understand the meaning from the authors.

4) In table 7, the model parameters are compared. They should also be compared with Yolos. 

The quality of English Language is fine.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper introduces a novel object detection model with a primary focus on traffic light classification from a distance. To achieve this goal, the authors propose an object model that combines pyramidal feature fusion and a self-attention mechanism. Experimental results demonstrate that the proposed method significantly enhances multiple performance metrics. Consequently, the method proves effective in traffic light recognition. In summary, the paper demonstrates strong structural and content motivation, with opportunities for further enhancement outlined below:

1. The abstract mentions the application of transformer-based techniques from NLP for improved outcomes. To enhance this claim, the authors could provide more compelling reasons for the adoption of these techniques.

2. Section 1 is relatively brief. Expanding on the study motivations, similar to the enriched content in the abstract, would bolster the section's depth.

3. While Section 2 presents an inclusive literature review, the selection of cited references may not completely align with the central theme. Employing more recent references to underscore study motivations could strengthen this section.

4. The paper employs numerous abbreviations. A single table summarizing key abbreviations would enhance reader understanding.

5. In Table 3, the authors could consider incorporating additional YOLO algorithms for comparison purposes, such as enhanced versions like YOLO5-YOLO8. Additionally, explaining the rationale behind the experimental design within the same section would provide clarity.

6. Figures presented could benefit from increased resolution for improved visibility. Furthermore, enriching the conclusions with insights into emerging trends in image analysis and understanding would add value to the overall discussion.

Minor editing of English language required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript presents a novel approach to traffic light detection using a combination of pyramidal feature fusion and self-attention mechanisms. The authors propose an object detection model that focuses on traffic light classification at a distance. The paper addresses the transition from CNN-based techniques to transformer-based techniques in this domain and proposes a model that combines the Parallel Residual Bi-Fusion (PRB) feature pyramid network, attention modules, architectural tuning, and optimizer selection. Experimental results demonstrate improvements in performance indicators, leading to notable enhancement in traffic light recognition.

 

Traffic light detection is an important aspect of computer vision, particularly in autonomous driving systems. The manuscript addresses this relevance and the potential improvements through the proposed model. The manuscript effectively identifies the challenges in traffic light detection, especially at a distance, and highlights the transition from CNN-based techniques to transformer-based techniques, providing a context for the proposed approach.

 

The manuscript outlines the proposed approach, which combines pyramidal feature fusion, self-attention mechanisms, and other architectural enhancements. The description of these components and their integration is clear and comprehensible. The manuscript presents experimental results that indicate improvements in performance indicators, although specific metrics are not mentioned. The use of experimental evidence strengthens the credibility of the proposed approach.

 

While the manuscript introduces the Parallel Residual Bi-Fusion (PRB) and attention modules, providing more technical details or references for these components would enhance the reader's understanding of their significance. Elaborate more on how the proposed pyramidal feature fusion, self-attention mechanisms, and other architectural enhancements work together to achieve improved traffic light detection. Besides, it is suggested to introduce more relevant work on object detection, such as such as: doi.org/10.3390/app122110808 and doi.org/10.1007/s00170-022-10335-8.

 

Provide specific details about which performance metrics were improved and by how much. Quantitative information would give a clearer understanding of the advancements achieved. Discuss how the proposed approach compares to existing techniques, especially in the context of transformer-based methods. This would provide a broader perspective on the approach's contribution.

 

In conclusion, the manuscript presents a valuable and well-written contribution to the field of traffic light detection. The proposed approach demonstrates relevance and innovation. Overall, I recommend the manuscript for acceptance with minor revision.

The quality of English in the manuscript is generally good. The writing is clear and effectively conveys the purpose, methodology, and findings of the proposed traffic light detection model integrating feature fusion and attention mechanism.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

This proposed problem has been solved in the revision. 

The English Language is fine.