Laser Radar Data Registration Algorithm Based on DBSCAN Clustering

Round 1

Reviewer 1 Report

The paper addresses and interesting and timely topic. Please find below a short list of comments that can improve the quality of the manuscript:

1) In the Introduction, there is only a very partial mention of the promising sensing capabilities of LIDAR sensors for target tracking. In the last years, indeed, there is a big interest from both academia and industry in trying to understand how lidar radars, but also more generally radars (e.g., FMCW) can be used to infer accurate position information of targets modeled as extended objects in space. This gives rise to the important problem of estimating the target contours, as well as associating the backscattered echo signals (also called point cloud registration). In this respect, the beginning of the Introduction should be enriched with some pointers to recent and relevant work on the topic, that clearly demonstrate the advantages of using radar sensors to accurately perform positioning and mapping. Prominent works that could be added are:

- "Cramér-Rao bound analysis of radars for extended vehicular targets with known and unknown shape", IEEE Transactions on Signal Processing, 2022.

- "Seamless tracking of apparent point and extended targets using Gaussian process PMHT", IEEE Transactions on Signal Processing, 2019.

2) In this reviewer's opinion, at the end of the Introduction it could be useful to add a table summarizing the most closely-related works existing in the literature and to highlight their main pros and cons. After that, authors should better emphasize how the present contribution advances the state-of-the-art.

3) Section 3 "Description of the algorithm" could be enriched with a graphical scheme depicting the workflow of the proposed method.

4) If possible, please improve the quality of the figures as they seem to render with low resolution.

5) It would be useful to corroborate the performance results reported in Section 4 with some computational complexity analyses, so as to highlight potential trade-offs between achieved accuracy in estimation and computational cost.

Author Response

Dear reviewer:

We are very grateful to your comments for the manuscript. According with your advice, we tried our best to amend the relevant part and made some changes in the manuscript. These changes will not influence the content and framework of the paper. All of your questions were answered below. And here we list the changes and marked in yellow in revised paper.

We appreciate for Reviewers’ warm work earnestly, and hope that the correction will meet with approval. If you have any questions, please contact us. 

Once again, thank you very much for your comments and suggestions.

Yours Sincerely,

Lei Zhang

 

Concern # 1:

In the Introduction, there is only a very partial mention of the promising sensing capabilities of LIDAR sensors for target tracking. In the last years, indeed, there is a big interest from both academia and industry in trying to understand how lidar radars, but also more generally radars (e.g., FMCW) can be used to infer accurate position information of targets modeled as extended objects in space. This gives rise to the important problem of estimating the target contours, as well as associating the backscattered echo signals (also called point cloud registration). In this respect, the beginning of the Introduction should be enriched with some pointers to recent and relevant work on the topic, that clearly demonstrate the advantages of using radar sensors to accurately perform positioning and mapping. Prominent works that could be added are:

- "Cramér-Rao bound analysis of radars for extended vehicular targets with known and unknown shape", IEEE Transactions on Signal Processing, 2022.

- "Seamless tracking of apparent point and extended targets using Gaussian process PMHT", IEEE Transactions on Signal Processing, 2019.

Author action:

Page 1, Row 38-49:We supplement the recent related work of lidar.

 

Concern # 2:

In this reviewer's opinion, at the end of the Introduction it could be useful to add a table summarizing the most closely-related works existing in the literature and to highlight their main pros and cons. After that, authors should better emphasize how the present contribution advances the state-of-the-art.

Author action:

Page 3, Row 99:We add a table to analyze the advantages and disadvantages of the current point cloud registration algorithm.

 

Concern # 3:

Section 3 "Description of the algorithm" could be enriched with a graphical scheme depicting the workflow of the proposed method.

Author action:

Page 16, Row 500:We use the flow chart to enrich the algorithm description in Section 3.

 

Concern # 4:

If possible, please improve the quality of the figures as they seem to render with low resolution.

Author action:

We have improved the quality of full-text images so that they are presented in high resolution.

 

Concern # 5:

It would be useful to corroborate the performance results reported in Section 4 with some computational complexity analyses, so as to highlight potential trade-offs between achieved accuracy in estimation and computational cost.

Author action:

Page 20, Row 597-614:In order to increase the computational complexity, we have analyzed the mean square deviation, absolute error and relative error. In the actual environment, due to occlusion, movement and low coincidence rate, not all measuring points have one-to-one true points, so the evaluation standard of mean square deviation is more unreliable. The absolute error can directly reflect the size of the error, and the relative error can more accurately reflect the accuracy of the registration algorithm. Therefore, this paper uses these two errors in order to take advantage of their advantages and evaluate the registration effect of data registration algorithm more directly and accurately.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscritp entitle: Laser Radar Data Registration Algorithm Based On DBSCAN Clustering by Yiting Liu et al., uses the base that the lidar data registration algorithm is limited by the search correspondence, which is complex and unstable for line-based algorithms, and has a low data coincidence rate for point-based algorithms. Hence, this work proposes a new algorithm based on DBSCAN clustering, which avoids the direct search for the corresponding relationship between points or lines. The algorithm involves a radar data preprocessing stage, a radar data registration stage, and a key cluster data selection stage, which uses kernel density estimation and K-L divergence to evaluate the similarity of two frames of point cloud after registration. The proposed algorithm improves the robustness of the algorithm and suppresses the influence of outliers on the algorithm. The authors claim that the error between the registration result and the actual value is within 10%, and the accuracy is better than the ICP algorithm. 

The research is intriguing and appears to be scientifically valid. However, before recommending its publication, it is important to take into account the following points:

Major 

While a laser radar data registration algorithm based on DBSCAN clustering has some advantages, there are also potential issues not consider properly in the present work:

1. One issue with DBSCAN is that it requires setting two parameters: the neighborhood radius (epsilon) and the minimum number of points required to form a cluster (minPts). Choosing appropriate values for these parameters can be challenging and can have a significant impact on the clustering results. In the context of laser radar data registration, the choice of these parameters can affect the accuracy and completeness of the resulting registration.

2. Another potential issue with DBSCAN is that it assumes that the clusters are of similar size and density. However, in practice, the point clouds generated by a laser radar system may have clusters of varying sizes and densities, which can make it challenging to identify corresponding clusters between point clouds.

3. In addition, DBSCAN may not be well-suited for handling highly complex or irregularly shaped clusters, which may occur in some laser radar data. In these cases, other clustering algorithms or registration techniques may be more appropriate.

 

4. Lastly, DBSCAN is a computationally intensive algorithm and may be slow to process large point clouds, which can limit its practical application in some scenarios.

Minor

1. The quality of Figures 2-16 is inadequate and the labels are not clear, which complicates the interpretation and analysis of the data. Some Figs appear framed and other ones not, etc. 

2. Please review the grammar, spelling, typos, plurals, and other potential errors in the document. Consider using the Grammarly tool to assist with a preliminary revision of the manuscript.

3. Certain references do not include a DOI, and some journal names are written in all capital letters. Although these issues can be addressed during the production phase, we suggest that authors follow the Author Guide on the website or template to standardize and regulate these details.

4. Conclusions are not supported with numerical values 

 

   

Author Response

Dear reviewer:

We are very grateful to your comments for the manuscript. According with your advice, we tried our best to amend the relevant part and made some changes in the manuscript. These changes will not influence the content and framework of the paper. All of your questions were answered below. And here we list the changes and marked in yellow in revised paper.

We appreciate for Reviewers’ warm work earnestly, and hope that the correction will meet with approval. If you have any questions, please contact us. 

Once again, thank you very much for your comments and suggestions.

Yours Sincerely,

Lei Zhang

 

Concern # 1:

One issue with DBSCAN is that it requires setting two parameters: the neighborhood radius (epsilon) and the minimum number of points required to form a cluster (minPts). Choosing appropriate values for these parameters can be challenging and can have a significant impact on the clustering results. In the context of laser radar data registration, the choice of these parameters can affect the accuracy and completeness of the resulting registration.

Author action:

Page 9, Row 308-312: For the selection of DBSCAN clustering algorithm parameters, this paper adopts the idea of k-dist graph proposed in the reference [1]. Calculate the distance from each point to the point nearest to its k, and sort these distances from large to small for drawing. The distance to find the inflection point is the value of, and the value of minPts is k+1.

• Yu, Y.;Zhou, A. An improved DBSCAN density algorithm. Computer Technology and Development. 2011, 21, 30-33. https://doi.org/10.3969/j.issn.1673-629X.2011.02.008.

 

Concern # 2:

Another potential issue with DBSCAN is that it assumes that the clusters are of similar size and density. However, in practice, the point clouds generated by a laser radar system may have clusters of varying sizes and densities, which can make it challenging to identify corresponding clusters between point clouds.

Author action:

In this paper, the two frames of point clouds are extracted into straight lines, and the straight lines of the two frames of point clouds are subtracted one by one. The DBSCAN algorithm is used to find the key clusters of rotation angle and translation matrix. That is, when the two straight lines are approximately parallel, their clustering results will form a concentrated cluster, and when the two straight lines are not parallel, their clustering results cannot form a very concentrated cluster, thus obtaining the rotation angle. The estimation of translation matrix is the same.Thanks for your good advice.

 

 

 

 

Concern # 3:

In addition, DBSCAN may not be well-suited for handling highly complex or irregularly shaped clusters, which may occur in some laser radar data. In these cases, other clustering algorithms or registration techniques may be more appropriate.

Author action:

In the algorithm of this paper, the key cluster formed by DBSCAN clustering is the cluster with the largest number of samples. That is, when the two lines are corresponding, the angle obtained by slope subtraction will appear in the key cluster. When the two lines are not corresponding, the angle distribution obtained by the slope subtraction is messy and may appear in other clusters with smaller samples. Therefore, for complex environments, some point clouds overlap, and the angles obtained by subtracting the slopes of the corresponding straight line segments of the two frames can be clustered in a small range and a key cluster with a large number of samples. Thank you for your good suggestions. We are also studying other clustering algorithms to find more suitable clustering algorithms for data registration and obstacle perception.

 

Concern # 4:

Lastly, DBSCAN is a computationally intensive algorithm and may be slow to process large point clouds, which can limit its practical application in some scenarios.

Author action:

Page 22, Row 657-658:This paper mainly deals with data registration of two-dimensional laser radar. Compared with three-dimensional laser radar, the amount of data is very small. And the amount of point cloud data can be reduced by adjusting the angle resolution and scanning angle of the lidar. Our next research content is to find more efficient clustering algorithms when dealing with large point clouds of 3D laser radar. Thank you for your suggestion.

 

Concern # 5:

The quality of Figures 2-16 is inadequate and the labels are not clear, which complicates the interpretation and analysis of the data. Some Figs appear framed and other ones not, etc.

Author action:

We have improved the quality of full-text images so that they are presented in high resolution.

 

Concern # 6:

Please review the grammar, spelling, typos, plurals, and other potential errors in the document. Consider using the Grammarly tool to assist with a preliminary revision of the manuscript.

Author action:

We revised the grammar, spelling, typos, plurals and other potential errors in the article and marked them in yellow in the pdf file.

 

 

 

 

Concern # 7:

Certain references do not include a DOI, and some journal names are written in all capital letters. Although these issues can be addressed during the production phase, we suggest that authors follow the Author Guide on the website or template to standardize and regulate these details.

Author action:

Page 22, Row 693-694:The DOI of the original reference was not found in the paper and the database, so we replaced a reference and supplemented the DOI.

Page 22, Row 700:We changed the name of the journal here to lowercase.

 

Concern # 8:

Conclusions are not supported with numerical values.

Author action:

Page 20, Row 617-627:We delete "when the position and attitude of the laser radar is large", and use absolute error and relative error to compare these two algorithms, so that the numerical value supports the conclusion.

Author Response File: Author Response.pdf

Reviewer 3 Report

The scanning of real objects using LIDAR is currently one of the modern technologies. The weak point of this acquisition of spatial data is its evaluation. So far, it has not been possible to find an efficient algorithm that can satisfactorily search for objects in a point cloud. Any solution proposal in this area is therefore very valuable. The authors propose an interesting algorithm based on clustering. The idea is not entirely original, but in this context it is, in my opinion, an innovative solution. I have no major comments on the article, only recommendations for improvement. Comments:

For section 2, the title "Method of solution" would be more appropriate.

The description of the algorithm in section 3 would be more appropriate written in a pseudo language (pseudo code) or a flow chart diagram.

In the Conclusions section, the authors compare their algorithm with the ICP algorithm. For the professional public, it would be advisable to supplement the comparison with other algorithms used to find efeatures in LIDAR data.

Author Response

Dear reviewer:

We are very grateful to your comments for the manuscript. According with your advice, we tried our best to amend the relevant part and made some changes in the manuscript. These changes will not influence the content and framework of the paper. All of your questions were answered below. And here we list the changes and marked in yellow in revised paper.

We appreciate for Reviewers’ warm work earnestly, and hope that the correction will meet with approval. If you have any questions, please contact us. 

Once again, thank you very much for your comments and suggestions.

Yours Sincerely,

Lei Zhang

 

Concern # 1:

For section 2, the title "Method of solution" would be more appropriate.

Author action:

Page 3, Row 123:We modify the title to "Method of solution".

 

Concern # 2:

The description of the algorithm in section 3 would be more appropriate written in a pseudo language (pseudo code) or a flow chart diagram.

Author action:

Page 16, Row 500:We use the flow chart to enrich the algorithm description in Section 3.

 

Concern # 3:

In the Conclusions section, the authors compare their algorithm with the ICP algorithm. For the professional public, it would be advisable to supplement the comparison with other algorithms used to find efeatures in LIDAR data.

Author action:

Page 20, Row 597-617:Thank you for your suggestion, because our experimental platform uses the ICP algorithm for data registration, and the ICP algorithm is a very mainstream data registration algorithm, many algorithms have been improved on the basis of it, so this algorithm is compared with the ICP algorithm. In this paper, absolute error and relative error are added to evaluate the accuracy of the two algorithms. Absolute error can directly reflect the size of the error, and relative error can more accurately reflect the accuracy of the registration algorithm. In this paper, these two errors are used to give full play to their advantages and evaluate the registration effect of data registration algorithm more directly and accurately.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors correctly addressed all my comments.

Reviewer 2 Report

Authors have considered  point-by-point suggestions and comments.

I recommend the publication of the revised version.