A 3D Lidar SLAM System Based on Semantic Segmentation for Rubber-Tapping Robot
Round 1
Reviewer 1 Report
This paper purpose a new 3D lidar SLAM system working in rubber plantation environment. Using semantic segmentation methods, the Lidar system filters out rubber tree trunk using as a localization reference, which also benefits the data storage and the speed on searching algorithm. The authors validated the slam system performance using their real data, which has better result compared to LOAM and LeGO-LOAM.
Several problems suggest to edit:
This paper doesn't show its error in orientation.
The author does not show the actual trajectory of recorded data, which didn't prove whether the mapping system can achieve loop closure.
Figure 8 shows that the trajectory distance is not really long. also missing comparison on z axis.
The performance of Lego-LOAM is relatively close to the purpose method. The error in figure 8 might caused by the unsuccessful filter on IMU, which can be proved by adding content on point 3.
The writing can be improved.
Author Response
Dear Reviewer,
We quite appreciate your favorite consideration and insightful comments. Now we have revised our manuscript entitled “A 3D Lidar SLAM System Based on Semantic Segmentation for Rubber-Tapping Robot” (ID: forests-2548909) exactly according to your comments, and found these comments very helpful. we hope this revision can make our paper more acceptable. The main corrections in the paper and the responses to your comments are as follows:
Point 1: This paper doesn't show its error in orientation.
Response 1: To display the error in orientation, we changed the index of the pose estimation error. The RMSE that originally contained only x and z axis information was replaced with the APE and RPE that included the information of timestamp, position (x, y, z), and orientation (x, y, z, w). Experimental results have been added to Table 1, Table 2 and Figure 9. And the corresponding description is given in the text.
Point 2: The author does not show the actual trajectory of recorded data, which didn't prove whether the mapping system can achieve loop closure.
Response 2: We added the image of the real scene, as shown in Figure 8a. The algorithms are compared with the recorded ground truth values in Figure 8b.
Point 3: Figure 8 shows that the trajectory distance is not really long. also missing comparison on z axis.
Response 3: We have increased the actual distance of Figure 8, as well as added the trajectory information on z axis. More details can be found in Figure 8b.
Point 4: The performance of Lego-LOAM is relatively close to the purpose method. The error in figure 8 might caused by the unsuccessful filter on IMU, which can be proved by adding content on point 3.
Response 4: We have not yet used the IMU in the LeGO-LOAM algorithm, and we have updated the trajectory in Figure 8. We hope it can meet your suggestion.
Special thanks to you for your good comments.
We tried our best to improve the manuscript and made some changes in the manuscript. These changes will not influence the content and framework of the paper. And here we did not list the changes but used the “Track Changes” function of MS Word.
We appreciate your and the Editors’ warm work earnestly and hope that the correction will meet with approval. Once again, thank you very much for your comments and suggestions.
Kind regards,
Hui Yang
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper proposes a semantic segmentation based SLAM method for rubber-tapping robot.
Questions:
- I feel the entire method pipeline is like a combination of paper RangeNet++ [23] and LOAM, since most key steps like Lidar point cloud to range images, semantic segmentation framework and so on come from [23] without change. What are the differences between these steps from this paper and [23]'s? If this is true, the authors must introduce [23] with more details and compare with it in thorough, instead of only citing it there.
- The SLAM method in this paper is based on a trunk point cloud segmentation step at first, but why can an accurate detection and reconstruction of trunks help improve the accuracy of SLAM trajectory, considering there are many other objects in the forest? Meanwhile, the paper also gives one further step to fit trunk point cloud to cylinders. Why does this step help SLAM too? Some ablation study must be added on these new and special steps.
- More details are needed about training data collection and annotation, since the trunk data used in this method is newly scanned on a new environment. What is the scope of this data? And how generally does it look like? More global and local pictures are better. Also, the annotation tool is interesting and also important to demonstrate the training data quality, please introduce more details about it.
- Again, about testing dataset seq1, seq2, seq3, better to have more pictures and details on them since they are new to readers.
- Figure 3b is not good enough to get many information from it. Either improve it with more parameters there, or delete it, since I think it's also from paper [23].
- Figure 6 is fine but not easy to see the big difference between methods. For instance, I feel the LeGO-LOAM result is at least on par with Se-LOAM. Better to give some specific comparison and/or new pictures from other viewpoints.
Not really.
Author Response
Dear Reviewer,
We quite appreciate your favorite consideration and insightful comments. Now we have revised our manuscript entitled “A 3D Lidar SLAM System Based on Semantic Segmentation for Rubber-Tapping Robot” (ID: forests-2548909) exactly according to your comments, and found these comments very helpful. we hope this revision can make our paper more acceptable. The main corrections in the paper and the responses to your comments are as follows:
Point 1: I feel the entire method pipeline is like a combination of paper RangeNet++ [23] and LOAM, since most key steps like Lidar point cloud to range images, semantic segmentation framework and so on come from [23] without change. What are the differences between these steps from this paper and [23]'s? If this is true, the authors must introduce [23] with more details and compare with it in thorough, instead of only citing it there.
Response 1: Firstly, the system in this paper added a semantic segmentation algorithm as a pre-processing to extract the trunk point clouds. The step in the section of the semantic section is from RangeNet++ and has not been changed yet. More important is the application in the background of rubber operation.
Secondly, it is about the difference between Se-LOAM and LOAM at the back end. If only added RangeNet++ to the front end of LOAM to extract the trunk features, other point cloud features in the original point clouds would be reduced. However, LOAM estimates the pose state by matching the nearest neighbor of the edge points and the plane points in adjacent frames. The reduction of the number of feature points will undoubtedly reduce the accuracy of pose estimation, as shown in Table 1 and Table 2 (LOAM and ‘LOAM + RangeNet++’).
Thus, instead of lidar odometry method in LOAM, Se-LOAM clusters the point cloud of tree trunks after semantic segmentation and simulates the synthesis of cylinders with the same radius and position to participate in feature matching. The pose transformation is calculated by searching the nearest neighbor points between the current point cloud and the corresponding cylindrical landmark in the projected frame. At each subsequent sweep, we combine the odometry output with landmarks for lidar mapping. For each tree feature in the odometry output, we assign it to a tree cylinder in the global world.
Therefore, back to the innovation of this paper. It is the first application of semantic SLAM for rubber-tapping robots. In addition, we changed the matching method of pose estimation in Lidar Odometry and Lidar Mapping. Compared with the ‘LOAM + RangeNet++’ algorithm, Se-LOAM achieves the function of feature extraction without losing the accuracy of pose estimation.
Point 2: The SLAM method in this paper is based on a trunk point cloud segmentation step at first, but why can an accurate detection and reconstruction of trunks help improve the accuracy of SLAM trajectory, considering there are many other objects in the forest? Meanwhile, the paper also gives one further step to fit trunk point cloud to cylinders. Why does this step help SLAM too? Some ablation study must be added on these new and special steps.
Response2: Thanks for the suggestion of ablation study. The trajectory and pose estimation comparison between ‘LOAM + RangeNet++’ algorithm and Se-LOAM algorithm have been added to Figure 8, Table 1, Table 2 and Figure9.
Simply combining RangeNet++ with LOAM will not improve the accuracy of SLAM trajectory. Due to the reduction of feature points, part of feature matching will be lost, and the precision of pose estimation will be reduced. Therefore, this paper changes the classical LOAM algorithm in Lidar Odometry and lidar Mapping, and does not use simple nearest neighbor matching of edge points and plane points. It participates in feature matching by clustering, fitting the point cloud of the tree trunk as a cylinder. The method is similar to Response 1.
Point 3: More details are needed about training data collection and annotation, since the trunk data used in this method is newly scanned on a new environment. What is the scope of this data? And how generally does it look like? More global and local pictures are better. Also, the annotation tool is interesting and also important to demonstrate the training data quality, please introduce more details about it.
Response 3: We have added a global map of the collected data set for better understanding to the reader. In addition, we changed the picture of the point cloud annotation tool with more details, as shown in Figure 3c. And a more specific text description is given.
Point 4: Again, about testing dataset seq1, seq2, seq3, better to have more pictures and details on them since they are new to readers.
Response 4: The trajectory of Figure 8b adds the real picture, as shown in Figure 8a. At the same time, we added real pictures to the three sets of sequences in Figure 7, as shown in Figure 7a.
Point 5: Figure 3b is not good enough to get many information from it. Either improve it with more parameters there, or delete it, since I think it's also from paper [23].
Response 5: Because the structure of the network part has not been changed in this article, the original network structure (Figure 3b in original manuscript) has been deleted.
Point 6: Figure 6 is fine but not easy to see the big difference between methods. For instance, I feel the LeGO-LOAM result is at least on par with Se-LOAM. Better to give some specific comparison and/or new pictures from other viewpoints.
Response 6: We have updated Figure 6 as your suggestions. Figure 6a is the real environment of the rubber plantations. Figure 6b is LOAM, whose point cloud is not clear. Figure 6c is the LeGO-LOAM algorithm. There are raised shrubs in this image, which can create problems for subsequent identification of tree trunks. Figure 6d is the Se-LOAM algorithm, which can only map the tree trunk point cloud.
Special thanks to you for your good comments.
We tried our best to improve the manuscript and made some changes in the manuscript. These changes will not influence the content and framework of the paper. And here we did not list the changes but used the “Track Changes” function of MS Word.
We appreciate your and the Editors’ warm work earnestly and hope that the correction will meet with approval. Once again, thank you very much for your comments and suggestions.
Kind regards,
Hui Yang
Author Response File: Author Response.docx
Reviewer 3 Report
Please indicate what are you trying to localize.
Please arrange the keywords alphabetically
In a separate paragraph indicate the importance of your work
The introduction is way too large. Please consider reducing the size of it.
You should dedicate a separate section for experiment setup rather than the results and discussion section.
Briefly, define what is the difference between Se-LOAM and other algorithms.
Please define what is Jetson TX2.
In figure 6, from what I see, LeGO-LOAM has better results, if you can add a real photo of the site, so the reader can understand more.
Fig 8 is not clear
In addition to what is presented in Figure 8, you should include the RMSE between every algorithm and the truth one, the result should be a number for each comparison, this will help the reader to understand the results more
Author Response
Dear Reviewer,
We quite appreciate your favorite consideration and insightful comments. Now we have revised our manuscript entitled “A 3D Lidar SLAM System Based on Semantic Segmentation for Rubber-Tapping Robot” (ID: forests-2548909) exactly according to your comments, and found these comments very helpful. we hope this revision can make our paper more acceptable. The main corrections in the paper and the responses to your comments are as follows:
Point 1: Please indicate what are you trying to localize.
Response 1: In this paper, a simultaneous localization and mapping (SLAM) system for rubber plantation is designed. The positioning refers to obtaining the attitude change of the robot by matching the key frame information, and obtaining the current pose estimation, to obtain the current positioning of the robot in the map.
Point 2: Please arrange the keywords alphabetically.
Response 2: Thank you for your suggestion. However, based on our search of published articles in this journal, most of the keywords are ranked in order of importance related to the article or appearance in the abstract, so we have not made changes to the keyword ranking.
Point 3: In a separate paragraph indicate the importance of your work.
Response 3: We added the contribution of the article in the penultimate paragraph of the profile to indicate the importance of our work.
Point 4: The introduction is way too large. Please consider reducing the size of it.
Response 4: We have made appropriate cuts to the introduction section. The order of introduction includes the background of rubber-tapping robot, the processing methods of SLAM front-end (traditional and network), the development and current shortcomings of SLAM, the research content and contribution points of this paper, and the summary of the remaining text.5.
Point 5: You should dedicate a separate section for experiment setup rather than the results and discussion section.
Response 5: We split the title of the Section 4. The experiment section is set as a separate Section 4. The discussion is split into a separate paragraph Section 5.
Point 6: Briefly, define what is the difference between Se-LOAM and other algorithms.
Response 6: The discussion explains how Se-LOAM differs from other algorithms is added to Section 5.
Point 7: Please define what is Jetson TX2.
Response 7: We replace the text with ‘This paper now describes a series of experiments to qualitatively and quantitatively analyze three SLAM methods, LOAM, LeGO-LOAM, and Se-LOAM, on the embedded system called Jetson TX2.’
Point 8: In Figure 6, from what I see, LeGO-LOAM has better results, if you can add a real photo of the site, so the reader can understand more.
Response 8: We have updated Figure 6 as your suggestions. Figure 6a is the real environment of the rubber plantations. Figure 6b is LOAM, whose point cloud is not clear. Figure 6c is the LeGO-LOAM algorithm. There are raised shrubs in this image, which can create problems for subsequent identification of tree trunks. Figure 6d is the Se-LOAM algorithm, which can only map the tree trunk point cloud.
Point 9: Fig 8 is not clear
Response 9: We replaced Fig 8 and added a picture of the real environment, as well as trajectory information on z axis.
Point 10: In addition to what is presented in Figure 8, you should include the RMSE between every algorithm and the truth one, the result should be a number for each comparison, this will help the reader to understand the results more
Response 10: Because the previous RMSE index only contained the information on x and y axis, we changed the index commonly used in the SLAM system, including APE and RPE. The pose errors of each algorithm are specifically written out in Table Table2. This advice allows the reader to intuitively understand the results. Thank you for your suggestion.
Specially thanks to you for your good comments.
We tried our best to improve the manuscript and made some changes in the manuscript. These changes will not influence the content and framework of the paper. And here we did not list the changes but used the “Track Changes” function of MS Word.
We appreciate your and the Editors’ warm work earnestly and hope that the correction will meet with approval. Once again, thank you very much for your comments and suggestions.
Kind regards,
Hui Yang
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
I think the authors' reply solves my concerns on the first version of the draft. I agree to accept it in present form.
