Large-Scale Outdoor SLAM Based on 2D Lidar

Round 1

Reviewer 1 Report

The authors propose a large-scale SLAM system using only 2D Lidar. The system is composed of three parts (front end, loop closure and back end) and the authors contribute to each part, mainly by modifying an existing algorithm to achieve a small performance improvement. All these contributions are well described and validated with several datasets.

It is always interesting to improve existing 2D Lidar SLAM methods, but nowadays there is a clear trend towards stereo or RGBD cameras and 3D Lidars. Those are better sensors and the technology advance is getting them cheaper. In fact, all the datasets used in this paper (even the ones made by the authors) have a 3D Lidar.

Regarding the results. Why table 1 shows only the comparison with Method 1? The authors claim that "during all the experiments, only two false loop closures are rejected". How many experiments with which datasets? Authors should be more specific.

It seems that loop closure detection and validation is only tested in campus #1 dataset. It would be interesting to see the comparison with the state-of-the-art methods in other datasets.

The authors claim that they tested their 2D-SLAM system on several other datasets and got good results. There is no need on explaining every dataset, but a table summarizing the main results would definitely validate the system.

One interesting thing to know is if your datasets are available for public use.

Some comments regarding paper presentation:

 - Figures and tables should appear after they are cited in the text, not before. Specifically, figures 1, 2, 5, 6, 8, 11 and 16 and table 1. It is an extra work for the reader to find out what is showing the figure or table before being described and cited in the text.

 - I don't understand the figures order in section 4.3. Figure 14 should go after figure 15.

 - The concept of AdaBoost is not introduced in the text. It is used several times without explaining what it is. Many readers will appreciate a simple explanation about this algorithm.

- Although the paper is well written in general, it should be revised for spell checking (e.g. "an wheel encoder" in line 239 should be "a wheel encoder"). Also, the authors should re-check which names have the different methods presented in the paper (e.g. "method 1" is sometimes "method #1") and try to be consistent.

Author Response

Response to Reviewer 1 Comments

Point 1: It is always interesting to improve existing 2D Lidar SLAM methods, but nowadays there is a clear trend towards stereo or RGBD cameras and 3D Lidars. Those are better sensors and the technology advance is getting them cheaper. In fact, all the datasets used in this paper (even the ones made by the authors) have a 3D Lidar.

Response 1: Thanks for your suggestions. 

Firstly, Compared with stereo or RGBD cameras,  lidar sensors have the advantages of high precision and strong resistance to interference, and is robust to illumination variations, especially in large-scale outdoor environment. 

Secondly, although 3D lidars are widely used in the field of UGV navigation, they work well in the mode of off-line mapping and on-line localization. SLAM based on 3D lidars seems to be unpractical in large-scale outdoor environment due to the large amount of computation and unreliability of loop closure detection, especially without GPS.

In this paper, we focus on the ability of practical application in large-scale outdoor environment. We believe our work makes contributions to SLAM based on 2D lidar. And these contributions can be applied into the framework of 3D lidar SLAM, which will be researched in our following works.

Point 2: Regarding the results. Why table 1 shows only the comparison with Method 1? The authors claim that "during all the experiments, only two false loop closures are rejected". How many experiments with which datasets? Authors should be more specific.

Response 2: Thanks for your reminder. 

Firstly, method 1 is the most similar one to our method, and the main difference is that we utilize more features as described in Sec. 3.2. So we focus on the improvement of our method compared with method 1.

Secondly, with regard to the performances of different loop closure detection methods, our first concern is the accuracy of detection. Because method 2 and method 3 obtain low detection accuracy, we think it is unnecessary to involve them in the comparison of time consumption.

Thirdly, We have modified Line 270-271 to be more specific. “During all the experiments including fifteen datasets from campus #1, one dataset from the Zhongdian Science and Technology Park and eleven datasets from KITTI benchmark, only two false loop closures are rejected ...”

Point 3: It seems that loop closure detection and validation is only tested in campus #1 dataset. It would be interesting to see the comparison with the state-of-the-art methods in other datasets.

Response 3: Thanks for your suggestions. 

Firstly, campus #1 dataset is representative with a lot of loop closures in large-scale environment. We believe that the comparison in this dataset is convincing. 

Secondly, other datasets such as the KITTI datasets only contains a small amount of loop closures. As explained in Line 263-264, “The subsequent experiments will show that our system can achieve nearly 0% false alarm for different types of datasets.” In the part of loop closure detection and validation, other datasets are mainly used for the further verification of the good performance of our method.

Point 4: The authors claim that they tested their 2D-SLAM system on several other datasets and got good results. There is no need on explaining every dataset, but a table summarizing the main results would definitely validate the system.

Response 4: Thanks for your suggestions. Firstly, we think the comparison of mapping details and paths are necessary to make the conclusion convincing. Secondly, we think that a table is less intuitive.

Point 5: One interesting thing to know is if your datasets are available for public use.

Response 5: Unfortunately, we're not going to make our datasets available for public use. However, we will upload one dataset along with our revised version. 

(/dataset/ZD.2d)

Point 6: Some comments regarding paper presentation:

Response 6: Thanks for your reminder.We have carefully checked the whole manuscript and have corrected several errors.

1. All figures and tables have been modified to appear after they are cited in the text.

2. The figures order have been carefully checked and modified.

3. A simple explanation about AdaBoost is given in Line 185-191.

4. We have corrected several typo errors.

Author Response File: Author Response.pdf

Reviewer 2 Report

This is an interesting paper on simultaneous localization and mapping using a 2D lidar. The paper is well written and the presented content is quite significant. I found it quite easy to follow but I would like to recommend some changes to improve its current presentation.

1. Please provide a table with math notation. Otherwise, it might be hard to follow.

2. Can you please explain with more details what Figure 1 illustrates? You can also add pointers on the different labels to direct us to the algorithm/math equation.

3. I would like to see more evaluations with prior work.

4. What is the computational advantage of your method compared to prior work?

Author Response

Response to Reviewer 2 Comments

Point 1: Please provide a table with math notation. Otherwise, it might be hard to follow.

Response 1: Thanks for your suggestions. All the math notations used in our paper are well explained where they appear for the first time and most of them are commonly used in the field of SLAM.

Point 2: Can you please explain with more details what Figure 1 illustrates? You can also add pointers on the different labels to direct us to the algorithm/math equation.

Response 2: Thanks for your reminder. We have modified Line 90-91 by adding pointers to explain Figure 1. “As shown in Fig. 1, our full system mainly contains three parts: the front-end (Sec. 3.1), the loop closure detection part (Sec. 3.2) and the back-end (Sec. 3.3).”

Point 3: I would like to see more evaluations with prior work.

Response 3: Thanks for your suggestions. As reviewer 1 has described, all our contributions are well described and validated with several datasets. And we have evaluated our system on our datasets and the KITTI datasets compared with the state-of-the-art methods.

Point 4: What is the computational advantage of your method compared to prior work?

Response 4: Thanks for your suggestions. Table 1 shows the computational advantage of our method for loop closure detection compared to prior work. Line 285-287 shows the computational advantage of our method compared to cartographer. Line 306-307 shows the computational advantage of our method compared to LOAM[16].

Author Response File: Author Response.pdf

Reviewer 3 Report

In this work, the authors propose a fully automated 2D simultaneous localization and mapping (SLAM) system based on LiDAR working in large-scale outdoor environment. The Authors improve the accuracy and robustness of the scan matching module by using an improved Correlative Scan Matching (CSM) algorithm.

Comments to Authors:

1- This paper is interesting but you should carfeully review the language.

Author Response

Response to Reviewer 3 Comments

Point 1: This paper is interesting but you should carefully review the language

Response 1: Thanks for your reminder. This paper has be polished by the person fluent in English and the English writing of the entire article has been checked and some sentences have been modified.

Author Response File: Author Response.pdf