Static Map Construction Based on Dense Constraints and Graph Optimization
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsDear authors,
The study is scientifically sound in general. However, some parts should be presented more clearly. Below, you can find some suggestions that in my opinion will help improve the manuscript.
Line 16 – use just point cloud instead of laser point cloud
Line 22 – you must spell out the abbreviations the first time you use them, e.g. LiDAR, but in the whole manuscript – all the abbreviations, LeGOLOAM, FOV, etc.
Line 28 – laser point cloud? just point cloud, or laser scanning data
Line 41, 42 – what do you mean by solid-state radar?
Line 65 – please explain in more detail the iVox data structure
Fig.1 The titles of the modules are not in accordance with the text (description) in lines 96,97, also, a bigger figure will help in terms of readability
Line 102 – the vehicle is equipped with… Do you have a specific one? Then describe it in detail. If not, then it is not correct because a mobile mapping system consists of various sensors, not just the ones you mentioned
Line 113 – change the title – especially the word Systems
Line 123 – laser data? It should be laser scanning data or a point cloud, laser data can refer to a very wide range of data.
Line 130- Please, describe (briefly) the approach from [18] it is important
Line 131, 132 rage image, distance image - better use depth image
Line 141 - … collected during map construction… it should be collected during data acquisition
Line 160 – do not use coverage area but …overlap
Line 161 - please explain the part… Due to the minimum angular resolution?
Line 183 – what do you mean by laser odometer?
Fig 3 a bigger figure is needed
Line 217. Please explain the sentence …. the requirements form potential constraints….
Fig. 4 A bigger fig is needed
Line 252 - Please, explain in more detail the procedure
Section 3.4. the approach has to be described in more detail
A larger Fig 5 is needed
Line 292 – please change the title of section 4
Line 296 – please explain in more detail the Carla simulator
Line 305 – please explain the European distance, and how it is calculated
The experiment has to be described in more detail. It is very hard to follow the Section 4 without detailed elaboration of the experiments and the results.
Table 1 – explain the content also in the text
Line 359 - Please change the title
Figure 12 – a larger one is needed, add a legend to explain the colors
Line 374 – explain the two methods
Line 377 – Why are XOZ, XOY, and YOZ not just XY, XZ, and YZ planes?
Line 378 – explain the monochrome mode
Fig 13 to FIG 15- please add legends, to explain the colors, the colored ellipses, circles, etc.
Before the conclusion some discussion about the strong and weak parts of your approach is needed.
Author Response
Dear Associate Editor:
We are grateful to all the reviewers for their thorough reviews. In accordance with the constructive comments, the paper has been carefully and extensively revised, and the revised parts have been underlined and given a green background in the revised manuscript for your convenience. Specifically, we have dealt with every critical issue you mentioned. Detailed response to each comment is given below. Every question has been answered and all the constructive comments are reflected in the revised manuscript. We look forward to hearing from you.
Q1: Line 16 – use just point cloud instead of laser point cloud.
A: Thanks for your suggestion. We have standardized the terminology for point clouds and changed "laser point cloud" to "point cloud."
Q2: Line 22 – you must spell out the abbreviations the first time you use them, e.g. LiDAR, but in the whole manuscript – all the abbreviations, LeGOLOAM, FOV, etc.
A: Thanks for your comprehensive suggestion. We have provided explanations for all the abbreviations in the paper.
Q3: Line 28 – laser point cloud? just point cloud, or laser scanning data.
A: Thanks for your suggestion. We have standardized the terminology for point clouds and changed "laser point cloud" to "point cloud."
Q4: Line 41, 42 – what do you mean by solid-state radar?
A: We are sorry for the spelling error caused by our carelessness. In the revised version, we have changed "solid-state radar" to "solid state LiDAR."
Q5: Line 65 – please explain in more detail the iVox data structure.
A: Thank you for your comment. The iVox (incremental voxels) data structure is modified from the traditional voxels and supports incremental insertion and parallel approximated k-NN queries. We have added an explanation of the ivox data structure to the manuscript.
Q6: Fig.1 The titles of the modules are not in accordance with the text (description) in lines 96,97, also, a bigger figure will help in terms of readability.
A: Thanks for your suggestion. We have enlarged the font in Figure 1 to improve readability. Additionally, we have revised the content in lines 128-129 to ensure consistency with the content in Figure 1.
Q7: Line 102 – the vehicle is equipped with… Do you have a specific one? Then describe it in detail. If not, then it is not correct because a mobile mapping system consists of various sensors, not just the ones you mentioned.
A: Thank you for your comment. Indeed, a mapping vehicle is typically equipped with various sensors, not just the ones I mentioned earlier. In this paper, the map data collection vehicle is specifically configured in the simulation scenario based on the algorithm requirements. Since the mapping algorithm in this study uses only point cloud data and IMU data, the vehicle is equipped with only two types of sensors: LiDAR and IMU. We have added a description of the data collection vehicle to the paper.
Q8: Line 113 – change the title – especially the word Systems.
A: Thank you for your suggestion. We have changed "Algorithms and Systems" to "Algorithm Implementation."
Q9: Line 123 – laser data? It should be laser scanning data or a point cloud, laser data can refer to a very wide range of data..
A: Thank you for your suggestion. We have changed "laser data" to "a point cloud."
Q10: Line 130- Please, describe (briefly) the approach from [18] it is important.
A: Thank you for your suggestion. In this paper, we employ the method proposed in reference [18] for semantic segmentation of point clouds. This method leverages the characteristics of point clouds for segmentation and does not involve deep learning, resulting in faster computation, making it suitable for integration into mapping systems. The semantic segmentation results allow us to identify and label noise points, such as outliers or points that do not belong to any category. Subsequently, we apply conditional filtering to remove these points marked as noise based on the segmentation results. We provide a brief introduction to the working principle of this method in the paper, while more detailed descriptions can be found in the original text.
Q11: Line 131, 132 rage image, distance image - better use depth image.
A: Thank you for your suggestion. We have changed "distance image" to "depth image."
Q12: Line 141 - … collected during map construction… it should be collected during data acquisition.
A: We apologize for the expression error caused by our carelessness and have revised the original text based on your suggestions.
Q13: Line 160 – do not use coverage area but …overlap.
A: We apologize for the expression error caused. We have changed " coverage area " to " overlap area."
Q14: Line 161 - please explain the part… Due to the minimum angular resolution?
A: Thank you for your suggestion. The "minimum angular resolution" refers to the smallest angle that a LiDAR can distinguish while rotating. This parameter determines the angular range covered by the laser beam emitted by the LiDAR, which in turn affects the detail and accuracy of the point cloud data. Specifically, the smaller the minimum angular resolution, the more environmental details the LiDAR can capture, resulting in more precise scanning results. Conversely, a larger minimum angular resolution may lead to insufficient coverage of certain areas during the LiDAR's rotation, resulting in incomplete scans or missing information. Therefore, when constructing local maps, the fusion of multiple frames of point clouds can effectively compensate for the information loss caused by the minimum angular resolution and occlusion from the scanning perspective. We have added a description of the minimum angular resolution to the paper.
Q15: Line 183 – what do you mean by laser odometer?.
A: Thank you for your comment. The laser odometer is used in SLAM algorithms to estimate the vehicle's pose information. Specifically, the laser odometer combines the pose information from the previous moment and estimates the device's relative displacement and rotation by comparing the current point cloud with the previous frame's point cloud. Therefore, it can only guarantee high accuracy within a local area. If too many frames are used when constructing the local map based on the pose provided by the laser odometer, it may lead to blurriness or ghosting effects in the local map. We have added a description of the laser odometer to the paper.
Q16: Fig 3 a bigger figure is needed.
A: Thank you for your suggestion. We have modified Figure 3 to enhance its readability.
Q17: Line 217. Please explain the sentence …. the requirements form potential constraints….
A: Thank you for your comment. We have revised the sentence and provided a brief explanation of "potential constraints" in the paper. "Potential constraints" refer to the possible constraints established by the system between keyframes after obtaining local maps and keyframes. These constraints are used to connect different keyframes, forming a pose graph for subsequent global graph optimization. Specifically, the construction process of potential constraints involves filtering keyframe pairs that meet specific criteria to form potential constraints, which may be based on relationships between adjacent frames or connections between historical loop frames. In the local map registration phase, the system calculates the specific values of all potential constraints and sets conditions to determine whether to accept or discard a particular constraint. This means that not all established potential constraints will be included in the final pose graph; only those that meet the accuracy requirements will be accepted. In this way, the method presented in this paper utilizes potential constraints to construct sparse or dense pose graphs, with a dense pose graph indicating that the system has introduced more observational data, thereby providing higher accuracy and better robustness during global optimization.
Q18: Fig. 4 A bigger fig is needed.
A: Thank you for your suggestion. We have modified Figure 4 to enhance its readability.
Q19: Line 252 - Please, explain in more detail the procedure.
A: Thank you for your suggestion. We have provided an additional description of the construction process of loop constraints to make it easier to understand.
Q20: Section 3.4. the approach has to be described in more detail.
A: Thank you for your suggestion. We have provided a more detailed description of the method in Section 4.4.
Q21: A larger Fig 5 is needed.
A: Thank you for your suggestion. We have modified Figure 5 to enhance its readability.
Q22: Line 292 – please change the title of section 4.
A: Thank you for your suggestion. We have changed the title of original Section 4 (now it is Section 5) to "Experiments."
Q23: Line 296 – please explain in more detail the Carla simulator.
A: Thank you for your suggestion. We have provided an additional description of the Carla simulator.
Q24: Line 305 – please explain the European distance, and how it is calculated.
A: Thank you for your comment. Due to our oversight, we mistakenly wrote "European distance" instead of "Euclidean distance." This has now been corrected.
Q25: The experiment has to be described in more detail. It is very hard to follow the Section 4 without detailed elaboration of the experiments and the results.
A: We have enhanced the content of the experimental section, including an introduction to the simulator, explanations of the evaluation metrics, and the setup of the simulation and experimental scenarios. Additionally, the images have been modified.
Q26: Table 1 – explain the content also in the text.
A: Thank you for your suggestion. We have provided a detailed explanation of the three metrics used to evaluate ATE: MEAN, RMSE, and MAX.
MEAN: This value represents the average ATE across all frames or segments of the trajectory. It provides a general indication of the overall accuracy of the trajectory estimation, with lower values indicating better performance.
RMSE (Root Mean Square Error): RMSE is a statistical measure that calculates the square root of the average of the squared differences between the estimated and true trajectories. This metric emphasizes larger errors more than smaller ones, making it a useful indicator of the precision of the estimates. A lower RMSE signifies a more accurate trajectory estimation.
MAX: This value indicates the maximum ATE recorded during the evaluation. It highlights the worst-case error encountered, providing insight into the most significant discrepancies between the estimated and true trajectories. A lower MAX value is preferable, as it reflects fewer substantial errors.
Together, these metrics offer a comprehensive view of the trajectory estimation accuracy, allowing for informed comparisons between different methods or configurations
Q27: Line 359 - Please change the title.
A: Thank you for your suggestion. We have changed the title of original Section 5.2 from "Real Car Experiment" to "Real-world Scenario Experiment.".
Q28: Figure 12 – a larger one is needed, add a legend to explain the colors.
A: Thank you for your suggestion. We have modified Figure 12 to enhance its readability. We added a description of the color of Figure 12 in lines 592-596.
Q29: Line 374 – explain the two methods.
A: Thank you for your suggestion. We have provided a brief explanation of the two methods in the revised manuscript.
Q30: Line 377 – Why are XOZ, XOY, and YOZ not just XY, XZ, and YZ planes?
A: Thank you for your suggestion. XOZ, XOY, YOZ, these are the three fundamental planes in the space coordinate system, representing different spatial directions. XZ, XY, YZ, these are often used to describe the position of a point on a particular plane, such as the point (x, y) on the XZ plane, the point (x, z) on the XY plane, and the point (y, z) on the YZ plane. Here is a plane expression of three-dimensional space, so XOZ, XOY, YOZ is more appropriate.
Q31: Line 378 – explain the monochrome mode.
A: Thank you for your suggestion. The monochrome mode here refers to using a single color to represent the result trajectory of a certain method. In the XOY plane trajectory comparison plot, we used two distinct colors to represent the outputs of the two methods. This intuitive color coding allows observers to quickly identify and compare the trajectory performance of both methods on the same plane. We have added an explanation of the monochrome mode and its function in the revised manuscript.
Q32: Fig 13 to FIG 15- please add legends, to explain the colors, the colored ellipses, circles, etc.
A: Thank you for your suggestion. We have revised Figures 13, 14, and 15 to enhance their readability. Additionally, we have provided a detailed explanation of the color meanings in Figures 14 and 15 in lines 613-621 of the revised manuscript. Furthermore, the colored circles in Figures 13-15 are merely to highlight areas of interest and do not have any other special significance.
Q33: Before the conclusion some discussion about the strong and weak parts of your approach is needed.
A: Thank you for your suggestion. In lines 562-578 of the revised manuscript, we analyze the limitations of the proposed method from three aspects: Impact of Scene Complexity, Low Loop Closure Trigger Frequency, and Maximum Error Issue. For detailed content, please refer to the revised manuscript.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe figures in the manuscript contain unusual symbols, such as artifacts from line breaks, which must be removed for clarity. All figures should be redrawn to ensure they are clean and free of such distractions.
The background of Figure 5 needs to be clean and free of any non-essential elements to enhance its clarity and presentation quality.
The baseline methods mentioned in the tables should be clearly specified with appropriate references. This is crucial for establishing a reliable comparison and for the readers’ understanding of the comparative analysis.
The 'Mapping Time Consumption' column in the tables predominantly shows 'N/A', which prevents a meaningful comparison with other methods. It would be beneficial to include time consumption data from alternative methods to provide a comprehensive comparative analysis.
The configuration of sensors used in the real-vehicle experiments should be detailed more thoroughly. This includes types, models, and settings of the sensors, which are essential for replicating the study and understanding the experimental setup.
For real-world experiments, consider utilizing the NCLT Dataset as a benchmark. This dataset can provide a standardized measure for comparison and would add value to the evaluation section by aligning it with recognized benchmarks in the field.
Comments on the Quality of English LanguageThe English could be improved to more clearly express the research.
Author Response
Q1: The figures in the manuscript contain unusual symbols, such as artifacts from line breaks, which must be removed for clarity. All figures should be redrawn to ensure they are clean and free of such distractions.
A: Thank you for your suggestions. We have revised all the images in the manuscript to ensure they are easy to understand and visually pleasing.
Q2: The background of Figure 5 needs to be clean and free of any non-essential elements to enhance its clarity and presentation quality.
A: Thank you for your suggestions. We have revised Figure 5.
Q3: The baseline methods mentioned in the tables should be clearly specified with appropriate references. This is crucial for establishing a reliable comparison and for the readers’ understanding of the comparative analysis.
A: Thank you for your valuable suggestions. We primarily compared the proposed method with the FAST-LIO-SLAM method in both virtual simulation environments and real-world scenarios. We also provided a brief introduction to the principles of FAST-LIO-SLAM and cited relevant references. In the original manuscript, we only discussed the comparison methods in virtual simulation experiments and did not cover real-world scenarios. The revised manuscript addresses this shortcoming.
Q4: The 'Mapping Time Consumption' column in the tables predominantly shows 'N/A', which prevents a meaningful comparison with other methods. It would be beneficial to include time consumption data from alternative methods to provide a comprehensive comparative analysis.
A: Thank you for your suggestion. The method in this paper focuses on offline mapping. Our comparison method, FAST-LIO-SAM, is a real-time online mapping method, so we did not compare mapping times with it.
Q5: The configuration of sensors used in the real-vehicle experiments should be detailed more thoroughly. This includes types, models, and settings of the sensors, which are essential for replicating the study and understanding the experimental setup.
A: Thank you for your valuable suggestion. We have added detailed information about the LiDAR and IMU sensor settings in both simulated and real-world scenarios to lines 375-382 of the revised manuscript.
Q6: For real-world experiments, consider utilizing the NCLT Dataset as a benchmark. This dataset can provide a standardized measure for comparison and would add value to the evaluation section by aligning it with recognized benchmarks in the field.
A: Thank you for your suggestion. We conducted a quantitative analysis of our method compared to FAST-LIO-SAM in simulated scenarios and in the real world. Since the simulated scenarios provide absolute ground truth, which allows for an effective evaluation of the method's performance, we did not conduct comparative tests on some public datasets.
Q7: Comments on the Quality of English Language. The English could be improved to more clearly express the research.
A: Thank you for your suggestion. We have carefully reviewed and rewritten the manuscript to ensure precise language, rigorous argumentation, and clear structure, thereby enhancing the overall readability. This process not only makes the article easier to understand but also improves the effectiveness of information transmission, helping readers to better grasp and comprehend the core ideas and details.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsIn this paper, the authors propose an environmental point cloud map construction method based on dense constraints and pose map optimization, which proposes an optimized and improved method to further control and eliminate the accumulation of errors. However, I would like to comment on some aspects to improve the quality of the manuscript:
- The article needs a Related Work section.
- The figures must be larger; it is difficult to read the content.
- Figure 5 must be converted into an algorithm. You can refer to https://overleaf.com/learn/latex/Algorithms.
- Is the scenario real or synthetic?
- Where were the data obtained?
- Which simulator was used for the experimentation?
- Were the patterns of the obstacle locations (buildings, houses, etc.) obtained from real data or placed randomly?
- What is the precision represented by the map construction?
- What metrics were used to evaluate the experimentation?
- The conclusions need to be improved, and future work must be added.
In this paper, the authors propose an environmental point cloud map construction method based on dense constraints and pose map optimization, which proposes an optimized and improved method to further control and eliminate the accumulation of errors. However, I would like to comment on some aspects to improve the quality of the manuscript:
- The article needs a Related Work section.
- The figures must be larger; it is difficult to read the content.
- Figure 5 must be converted into an algorithm. You can refer to https://overleaf.com/learn/latex/Algorithms.
- Is the scenario real or synthetic?
- Where were the data obtained?
- Which simulator was used for the experimentation?
- Were the patterns of the obstacle locations (buildings, houses, etc.) obtained from real data or placed randomly?
- What is the precision represented by the map construction?
- What metrics were used to evaluate the experimentation?
- The conclusions need to be improved, and future work must be added.
Author Response
Q1: The article needs a Related Work section.
A: Thank you for your suggestion. We have specifically updated Section 2 of the revised manuscript to include a detailed summary of related works, focusing on the recent developments in laser point cloud mapping techniques.
Q2: The figures must be larger; it is difficult to read the content.
A: Thank you for your suggestions. We have revised all the images in the manuscript to ensure they are easy to understand and visually pleasing。
Q3: Figure 5 must be converted into an algorithm. You can refer to https://overleaf.com/learn/latex/Algorithms.
A: Thank you for your suggestion. We have revised the content in Figure 5 to be more detailed pseudocode.
Q4: Is the scenario real or synthetic?
A: Thank you for your comments. In Section 5.1, the virtual simulation experiments involve three different scenarios, each of which is artificially constructed using the CARLA software. In Section 5.2, which covers experiments in real-world scenarios, the settings are based on the location of our laboratory and the surrounding buildings.
Q5: Where were the data obtained?
A: Thank you for your review comments. In the CARLA simulation environment, we equipped a Tesla Model 3 with a Velodyne 32-line LiDAR and an IMU, collecting LiDAR point cloud data and IMU data via this simulated vehicle. In the real-world experiments, we equipped a drive-by-wire chassis vehicle with the same specifications of Velodyne 32-line LiDAR and a hipnuc-CH104M 9-axis IMU, collecting experimental data through this vehicle. We have added these experimental setups to the revised manuscript.
Q6: Which simulator was used for the experimentation?
A: Thank you for your review comments. In the virtual environment, we use the CARLA software to simulate autonomous vehicles. We have supplemented the revised manuscript with an introduction to CARLA and included the relevant references to cite CARLA. Additionally, the autonomous driving software platform is built using the Robot Operating System (ROS).
Q7: Were the patterns of the obstacle locations (buildings, houses, etc.) obtained from real data or placed randomly?
A: Thank you for your review comments. In the virtual scenario, the layout of buildings, trees, traffic facilities, static vehicles, and other objects is manually set in a random manner.
Q8: What is the precision represented by the map construction?
A: Thank you for your review comments. We assess the accuracy of mapping by comparing the trajectory differences after mapping completion. Since the virtual simulation world provides accurate ground truth, we leveraged this advantage to conduct a comparative analysis of the mapping accuracy of the method proposed in this study and the comparison method. As shown in Tables 1, 2, and 3, the maximum absolute trajectory difference of our method is only 0.16 meters, with an average maximum of 0.057 meters. Moreover, our method significantly outperforms the comparison method, FAST-LIO-SAM, across all evaluation metrics for absolute trajectory differences, demonstrating the superior performance of our approach.
Q9: What metrics were used to evaluate the experimentation?
A: Thank you for your review comments. In the virtual simulation experiments, we evaluate the absolute trajectory differences from three main aspects: mean, root mean square error (RMSE), and maximum (MAX). The specific data is as follows:
MEAN: This value represents the average ATE across all frames or segments of the trajectory. It provides a general indication of the overall accuracy of the trajectory estimation, with lower values indicating better performance.
RMSE (Root Mean Square Error): RMSE is a statistical measure that calculates the square root of the average of the squared differences between the estimated and true trajectories. This metric emphasizes larger errors more than smaller ones, making it a useful indicator of the precision of the estimates. A lower RMSE signifies a more accurate trajectory estimation.
MAX: This value indicates the maximum ATE recorded during the evaluation. It highlights the worst-case error encountered, providing insight into the most significant discrepancies between the estimated and true trajectories. A lower MAX value is preferable, as it reflects fewer substantial errors.
We have added the introduction of these three specific indicators to the revised manuscript.
In the real world, we primarily use visualization to qualitatively assess the performance differences between the two methods.
Q10: The conclusions need to be improved, and future work must be added.
A: Thank you for your suggestions. In the revised manuscript, lines 601-621 analyze the shortcomings of our method. Then, in the final conclusion section, we have added a perspective on future research, namely: In future studies, we plan to use deep learning techniques to perform more precise segmentation of point cloud data, effectively filtering out noise points and dynamic objects. This step will help reduce the impact of these factors on the accuracy of the mapping algorithms, thereby enhancing the robustness and adaptability of our approach. Additionally, we are considering applying deep learning techniques to loop closure detection, by aligning features more accurately, to further improve the precision of loop closure detection. These measures are expected to significantly enhance the overall performance of the system.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsPlease once again check the colors in the fig 15... the dots representing the baseline are also multicolored not just red as in the legend
Author Response
Thanks to your suggestion, we have changed Figures 14 and 15 to remove the legend from the figure and add a description of the starting point.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsNo more comments.
Comments on the Quality of English LanguageThe Quality of English is fine.
Author Response
Thank you for your excellent suggestions on this article.
Author Response File: 
Author Response.docx
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors are thanked for incorporating the suggested revisions. Nonetheless, a brief review of the English language would further enhance the clarity and precision of the article.
Comments on the Quality of English LanguageThe authors are thanked for incorporating the suggested revisions. Nonetheless, a brief review of the English language would further enhance the clarity and precision of the article.
Author Response
Thank you for your suggestion. We have reorganized the English expressions in the article, especially the descriptions of professional terms. Please review, thank you.
Author Response File: Author Response.pdf
