Multi-Source Heterogeneous Data Fusion Algorithm for Vessel Trajectories in Canal Scenarios

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper, the authors introduce the SSGDA Framework as a multi-source data fusion algorithm. The paper is well written; however, there are some minor comments:   1. Formatting of references in text looks faulty; it seems the references are in superscript. 2. Some references in Reference section are "censored"; there is ** instead of letters. Faulty references include references 1, 2, 3, 26, 32, 37. Double check all references. 3. In Section 3.1., you introduce Figure 2 with "as follows:", without a cross-reference; you should add a cross-reference here. 4. Figure 4 is not cross-referenced. Also, you should check whether its size is appropriate according to Electronics Journal rules or if it is outside of the margins. Either way, it should not be hard to rearrange image to fit the page margins. 5. In section 3.2 you started "a list" (1) Kalman filter (2) Feature Extraction (3) Hungarian algorithm, without introducing why it is there; paraphrase that. 6. Equations: a) You should consider splitting Equations 3 and 4; either as 3a, 3b, 4a, 4b or as 3, 4, 5, 6. b) Some Equations are not cross-referenced (7, 8, 9, 10, 16, 17, 18, 20, 21, 22) - double-check all Equations. c) Fix formatting in Equation 19. d) Some Equations should be listed as Equations, rather than inline text, i.e.: I)Kalman gain: 𝐾𝑘 = 𝑃𝑘|𝑘−1𝐻⊤(𝐻𝑃𝑘|𝑘−1𝐻⊤ + 𝑅)−1 II) appearance cost: 𝑎𝑖𝑗 = 1 − ⟨𝑓𝑖,𝑔𝑗⟩/∥𝑓𝑖∥∥𝑔𝑗∥ 7. Can you explain the "video trajectory - AIS trajectory" matching task as binary classification in more detail? Does it consider exact matches only, or is there some error margin? 8. Figure 7 and Figure 9 - it is unclear does they visualize accuracy or precision (accuracy in caption, precision on the y-axis) 9. Figure 10 - in the caption say explicitly that it vizualises MSE. 10. In line 572, I guess it should be Figures 7 and 8, instead Figures 1 and 2.

Author Response

Comments 1: Formatting of references in text looks faulty; it seems the references are in superscript.

 

Response 1: Many thanks for your useful comments.

I fully agree with your suggestion and have noticed an issue with the reference formatting in the paper, where the citation markers were incorrectly set to superscript format. In response, I have corrected the formatting throughout the entire manuscript and adjusted all references to adhere to the standard format required by the journal.

 

Comments 2: Some references in Reference section are "censored"; there is ** instead of letters. Faulty references include references 1, 2, 3, 26, 32, 37. Double check all references.

 

Response 2: Many thanks for your useful comments.

In response to the issue of the "**" substitute characters in the references, I have thoroughly reviewed and completed the relevant citations to ensure they are correctly displayed and comply with the formatting requirements of the journal.

 

Comments 3: In Section 3.1., you introduce Figure 2 with "as follows:", without a cross-reference; you should add a cross-reference here.

 

Response 3: Many thanks for your useful comments.

Regarding the issue of missing graphic cross-references, I have added a note in line 315 of the revised manuscript to provide a clear explanation, ensuring a proper connection between the chart and the text. The modified text is as follows: Based on the preceding discussion, the raw AIS data undergoes a multi-stage prepro-cessing procedure, which is visualized in the flowchart of Figure 2.

 

Comments 4: Figure 4 is not cross-referenced. Also, you should check whether its size is appropriate according to Electronics Journal rules or if it is outside of the margins. Either way, it should not be hard to rearrange image to fit the page margins.

 

Response 4: Many thanks for your useful comments.

Regarding the issue of Figure 4 not being cross-referenced, I have added the appropriate cross-references in the text to establish a clear connection between the figure and the content. The addition can be found in line 410, with the revised text reading: "The overall tracking process is depicted in Figure 4." Additionally, I have resized Figure 4 to comply with the journal's formatting requirements and adjusted the image dimensions to meet the specifications for page margins.

 

Comments 5: In section 3.2 you started "a list" (1) Kalman filter (2) Feature Extraction (3) Hungarian algorithm, without introducing why it is there; paraphrase that.

 

Response 5: Many thanks for your useful comments.

In the revised draft submitted, I have added a detailed explanation of the specific functional advantages of the three modules—Kalman filter, feature extraction, and Hungarian algorithm—between lines 386 and 416. I also elaborated on their application in this study, as well as the rationale and role they play in addressing the problem. With these modifications, the content is now more coherent and easier to understand.

 

Comments 6: Equations: a) You should consider splitting Equations 3 and 4; either as 3a, 3b, 4a, 4b or as 3, 4, 5, 6. b) Some Equations are not cross-referenced (7, 8, 9, 10, 16, 17, 18, 20, 21, 22) - double-check all Equations. c) Fix formatting in Equation 19. d) Some Equations should be listed as Equations, rather than inline text, i.e.: I) Kalman gain:  II) appearance cost:    

 

Response 6: Many thanks for your useful comments.

1. a) Regarding Equations (3) and (4), I have split them into (3a), (3b), (4a), and (4b) and numbered them accordingly at line 358 to improve the clarity and readability of the formulas. Additionally, Equation (7) has also been split into (7a) and (7b) to address the same issue.
2. b) I have also revised the cross-references for the equations to enhance the readability and connection between the formulas and the text description.
3. c) On line 624, I have corrected the format of Equation (19), particularly the "min" function inside the formula. I added the separator "," to improve the clarity, i.e., , make it conform to the format requirements of the journal.
4. d) For certain formulas, I have changed them from inline display to the formal equation format, located at lines 447 and 485 of the revised manuscript, respectively. Specifically, I have separated the following formulas: the Kalman gain formula

as Equation (9), and the appearance cost formula

as Equation (12). The subsequent formulas have been renumbered accordingly. This change improves the readability and clarity of the formulas, ensuring compliance with the journal's formatting requirements.

 

Comments 7: Can you explain the "video trajectory - AIS trajectory" matching task as binary classification in more detail? Does it consider exact matches only, or is there some error margin?

 

Response 7: Many thanks for your useful comments.

We have added a detailed description of the matching mechanism in Section 3.3.2 (i.e., lines 559-573). Specifically, the matching of video tracks with AIS tracks is modeled as a binary classification problem, aiming to determine whether two tracks correspond to the same target. Given the differences in sampling frequency and coordinate accuracy between various data sources, we introduce a spatial tolerance threshold, denoted as . When the Euclidean distance between trajectory points is less than , the match is considered valid, and a positive score is assigned. If the distance exceeds , a penalty is applied. This approach allows the algorithm to tolerate a certain range of errors, enabling efficient trajectory matching and data correlation.

 

Comments 8: Figure 7 and Figure 9 - it is unclear does they visualize accuracy or precision (accuracy in caption, precision on the y-axis)

 

Response 8: Many thanks for your useful comments.

In response to the "precision" and "accuracy" issues in Figures 7 and 9 that you mentioned, we have standardized the descriptions in the charts and annotations, consistently using "precision." At the same time, we have also revised the relevant terminology throughout the entire manuscript to ensure consistency.

 

Comments 9: Figure 10 - in the caption say explicitly that it vizualises MSE.

 

Response 9: Many thanks for your useful comments.

I fully agree with your suggestion and have made improvements accordingly. Specifically, I have rephrased the caption for Figure 10, located on line 740 of the revised manuscript. The revised caption now reads: "Comparison of MSE and  between the baseline (without E-PSO) and the proposed method on the FVessel dataset."

 

Comments 10: In line 572, I guess it should be Figures 7 and 8, instead Figures 1 and 2.

 

Response 10: Many thanks for your useful comments.

There was indeed an issue with the figure numbering in the text. I have corrected it on line 697, and the revised text now reads: "Figures 7 and 8."

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Suggestions for improving the article:

– Include a flowchart or diagram illustrating the entire pipeline (AIS preprocessing, video extraction, E‑PSO, DBRP‑Match) to facilitate an overall understanding.

– Explain the rationale for selecting the values ω, c₁, and c₂ in E‑PSO and the tolerance δ in DBRP‑Match (e.g., preliminary studies or tuning on the validation set).

– Add a subsection that directly compares SSGDA with other multi‑source frameworks developed in the past 2–3 years, especially those integrating LiDAR or radar.

Discussion of practical applications
– Provide concrete usage examples (e.g., traffic monitoring in specific well‑known channels) and outline pilot plans for field validation.

Author Response

Comments 1: Include a flowchart or diagram illustrating the entire pipeline (AIS preprocessing, video extraction, E‑PSO, DBRP‑Match) to facilitate an overall understanding.

 

Response 1: Many thanks for your useful comments.

I completely agree with your suggestion. In the overall fusion flowchart—Figure 2 (SSGDA trajectory matching framework)—I have added detailed explanations covering AIS preprocessing, video extraction, E-PSO trajectory rotation and translation, DBRP-Match trajectory matching, etc. These additions are intended to help readers more intuitively understand the relationships and workflow between the components.

You can find these details in the newly submitted revised manuscript, between lines 268 and 298.

 

Comments 2: Explain the rationale for selecting the values , , and  in E‑PSO and the tolerance  in DBRP‑Match (e.g., preliminary studies or tuning on the validation set).

 

Response 2: Many thanks for your useful comments.

The parameters , , and  are set according to the default settings of the PSO algorithm, and their validity is confirmed through experimental verification. The threshold   in this paper is then optimized through multiple experiments on the dataset, with the optimal value being selected as the matching threshold. This ensures a good balance between preventing both overfitting and underfitting, and is used to determine whether trajectory points match. In future work, we will continue to conduct further experiments with these three parameters and the threshold, further adjusting and optimizing them to enable dynamic parameter tuning.

 

Comments 3: Add a subsection that directly compares SSGDA with other multi‑source frameworks developed in the past 2–3 years, especially those integrating LiDAR or radar.

 

Response 3: Many thanks for your useful comments.

In the revised manuscript, specifically in the "Related Work" section (lines 215 to 241), I have added a comparative analysis of various multi-source fusion frameworks. This section includes a comparison of fusion frameworks based on three different data source types from the past three years. I provide a detailed analysis of the advantages and disadvantages of these frameworks and propose future design frameworks to address the limitations of existing approaches. Moving forward, I plan to conduct further research on multi-source fusion frameworks to refine and optimize my methodology.

 

Comments 4: Provide concrete usage examples (e.g., traffic monitoring in specific well‑known channels) and outline pilot plans for field validation.

 

Response 4: Many thanks for your useful comments.

In line 788 of the revised manuscript, we have added relevant content. Specifically, this framework will be applied to the Pinglu Canal basin in the future, which is currently under construction. Cameras and AIS receivers have already been deployed around the basin, and field verification of the technology application is planned for the near future.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper proposes a novel data fusion and trajectory matching framework—SSGDA (Shape Similarity and Generalized Distance Adjusting)—for aligning vessel trajectories obtained from heterogeneous sources such as AIS and video surveillance in canal environments.

The contributions are clearly highlighted: an Enhanced Particle Swarm Optimization (E-PSO) algorithm was designed for rigid transformation and trajectory alignment; and a Distance-Based Reward-Penalty Matching (DBRP-Match) algorithm that balances local and global similarity was proposed.

The proposed solution is tested on the FVessel dataset and shows significant improvement in matching accuracy and MSE reduction compared to other existing methods.

The paper is technically sounds, it is well-structured and written, relevant and timely. It has an interesting topic, the gap in the field was well identified, and the proposed method is innovative.

The methodology is well described and solid, combining object detection (YOLOv11 + DeepSORT), spatiotemporal data alignment, and optimization techniques. The technical descriptions are detailed and well-supported by figures.

The state-of-the-art is well presented. Most references are relevant and within the last 5 years. No excessive self-citation.

The results are clearly presented. The conclusions are consistent.

The figures are appropriate, the properly show the data/results.

 

There are some minor observations:

• The existing methods used for comparison should have more details/explanations/description.
• At methods (E-PSO) - ε = 50 is arbitrarily chosen? — please justify. And this observations is valid for all the parameters..why did you choose those values?
•  Results - It would be interesting to add some results about the distance errors before/after.

Author Response

Comments 1: The existing methods used for comparison should have more details/explanations/description.

 

Response 1: Many thanks for your useful comments.

Regarding the descriptive issue of the "contrast method" that you raised, I fully agree with your suggestion. To provide readers with a clearer understanding of the comparative approaches we have adopted, I have included a more detailed comparative analysis of the advantages and disadvantages of these methods in lines 697-716 of the manuscript. This aims to better support the strengths of our proposed SSGDA framework. For a detailed overview of the revisions, please refer to the latest submitted manuscript.

 

Comments 2: At methods (E-PSO) -  is arbitrarily chosen? — please justify. And this observations is valid for all the parameters. why did you choose those values?

 

Response 2: Many thanks for your useful comments.

In response to this issue, I would like to explain as follows. Regarding the selection of  values, we have conducted experimental verification on the dataset. The results indicate that a threshold that is too low leads to insufficient trajectory translation and rotation, while a threshold that is too high causes overfitting and computational redundancy. Moving forward, we will conduct further experiments on this issue to adjust and optimize these parameters, aiming for dynamic parameter adjustment.

 

Comments 3: Results - It would be interesting to add some results about the distance errors before/after.

 

Response 3: Many thanks for your useful comments.

I fully agree with your suggestion to "add the results before and after the distance error" and believe it will help to demonstrate the effectiveness of our approach in a more comprehensive manner. Therefore, I have included the relevant data in lines 746-748 of the submitted manuscript. Specifically, I added the mean square error of trajectory translation and rotation without the E-PSO algorithm, as well as the mean square error after applying the E-PSO algorithm. I also compared the reduction in mean square error following the addition of the E-PSO algorithm. For detailed information, please refer to the latest submitted manuscript.

Author Response File: Author Response.pdf