Image-Based Obstacle Detection Methods for the Safe Navigation of Unmanned Vehicles: A Review

Round 1

Reviewer 1 Report

Overall, this paper is well written, the previous related methods were well summarized, and the future research suggestions were provided. The following papers investigate deep learning and monocular on depth estimation may be considered as the reference supplement.

“Stereo Object Matching Network.” 2021 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 12918–12924.

“Fast-PGMED: Fast and Dense Elevation Determination for Earthwork Using Drone and Deep Learning.” Journal of Construction Engineering and Management, 148(4), https://doi.org/10.1061/(ASCE)CO.1943-7862.0002256.

“Efficient Deep Learning for Stereo Matching.” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 5695–5703.

Author Response

Reviewer #1:

Overall, this paper is well written, the previous related methods were well summarized, and the future research suggestions were provided.

1. The following papers investigate deep learning and monocular on depth estimation may be considered as the reference supplement.

“Stereo Object Matching Network.” 2021 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 12918–12924.

“Fast-PGMED: Fast and Dense Elevation Determination for Earthwork Using Drone and Deep Learning.” Journal of Construction Engineering and Management, 148(4), https://doi.org/10.1061/(ASCE)CO.1943-7862.0002256.

“Efficient Deep Learning for Stereo Matching.” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 5695–5703.

Answer: Thank you. We appreciate the time and effort you put into reviewing our paper. To address this comment, the suggested papers were studied, and the relevant parts in our paper were revised. Also, the references were added to the manuscript as the reference supplement (line 51-Table 1, lines 574-578).

Author Response File: Author Response.docx

Reviewer 2 Report

The authors provided a comprehensive review with many relevant and recent references, which is considered as a plus for their study. However, review papers are normally expected to provide critique and authors’ personal perspective towards the investigated subject. Indeed, there are some useful points in the last section, but it would be preferable to add a section “Discussion and prospects” before the conclusions. On this context, conclusions appear as independent on the author’s views and simply gather recorded facts. Therefore the authors are asked to reorganize the structure of the paper at this point. Also, at the end of sections 3 and 4 it is recommended to put a small subsection “summary” to summarize the most important remarks, or advantages/limitations of each category. Finally, adding an abbreviation list at the end of the paper would be useful for the readers.

Author Response

Reviewer #2:

The authors provided a comprehensive review with many relevant and recent references, which is considered as a plus for their study.

1. However, review papers are normally expected to provide critique and authors’ personal perspective towards the investigated subject. Indeed, there are some useful points in the last section, but it would be preferable to add a section “Discussion and prospects” before the conclusions. On this context, conclusions appear as independent on the author’s views and simply gather recorded facts. Therefore the authors are asked to reorganize the structure of the paper at this point.

Answer: Thank you for your comment. Based on your comment, we went through the paper and added a number of discussions that represented our views on the presented techniques. These can be found in lines 227-231, lines 275-279, lines 376-380, lines 439-445, lines 519-523 and lines 660-663.

In addition, as suggested, we included a new section “5. Discussions and prospects” (lines 692-777), which in addition to giving an overview of the obstacle detection merits and disadvantages, presents a number of points we envisage as future trends or challenges to be resolved. Please note that in this section, Table 9 in the original paper which was in Conclusions, was moved to this new section.  

Also, some parts of the Conclusions (lines 808-811, lines 818-821 and lines 837-842) were revised to present more in-depth points and to suggest some lines of work to be pursued by other researchers. Also, a new subsection, i.e. Recommendations (section 6.1, lines 843-872), was added that summarises our suggestions for future developments and research.

2. Also, at the end of sections 3 and 4 it is recommended to put a small subsection “summary” to summarize the most important remarks, or advantages/limitations of each category.

Answer: Very good suggestion. For this purpose, subsections "3.5" and "4.3" were added to the manuscript (lines 447-482 and lines 673-691).

3. Finally, adding an abbreviation list at the end of the paper would be useful for the readers.

Answer: Thanks. A list of abbreviations and acronyms used in the paper was added (line 875-Table 12). 

Author Response File: Author Response.docx

Reviewer 3 Report

This paper reviews different image processing based methods for obstacle detection for unmanned vehicles. The paper is well written, easy to understand and give very good comparison and state of the art methods. I would recommend accepting in the present form. 

Author Response

Reviewer #3:

This paper reviews different image processing based methods for obstacle detection for unmanned vehicles. The paper is well written, easy to understand and give very good comparison and state of the art methods. I would recommend accepting in the present form. 

Answer: Thank you very much for the time and effort you spent reviewing our paper.

Author Response File: Author Response.docx

Reviewer 4 Report

General comments

The manuscript gathers context about different image-based obstacle detection methods. The paper first starts with the division of two categories: Stereo and Monocular. Then, elaborates it by describing the four techniques (Appearance-based, Motion-based, depth-based, and expansion-based) as part of the monocular category. Later, stereo-based approaches are divided into Inverse Perspective Mapping and Disparity histogram-based methods. The author then moves on to the abilities required from an obstacle detection algorithm such as narrow and small obstacle detection, moving obstacle detection, obstacle detection in all directions and fast/real-time obstacle detection. Several articles were cited, annotating their usage, principles, strengths and the criteria they met. Furthermore, the authors elaborated more on the four categories that comprehend the monocular method and continued to talk about the stereo-based method for terrestrial and aerial robots. The paper concludes by addressing the fact that Monocular-based obstacle detection approaches use only one camera. They are simple computations and are fast. It is noted that it was implemented on aerial and terrestrial robot navigations. Thus, it is mentioned that for Stereo-based methods there is a need for using of powerful Graphics Processing Units. Finally, it is stated that future methods will focus on detecting narrow and small obstacles and that deep learning techniques have helped reduce the computation time, for both monocular as well as stereo-based obstacle detection techniques.

Unfortunately, the reviewer feels the paper has not reached the level of review paper publication, the organization is poor, and not diverse enough;

1. The beginning of the paper introduces the Stereo and Monocular methods and later moved on to illustrate the four techniques used by Monocular-based methods, following the tree diagram in Figure1. However, when it is the turn of a stereo-based method, the paper focuses on the stereo-based obstacle detection methods for terrestrial and aerial robots (Instead of IPM-Based method and Disparity Based method).

2. The stereo portion does not align with the structure of the paper that the author has followed in the monocular portion previously, therefore, can cause confusion to the reader.

3. The reviewer believes this can appear as the stereo method section was not described as thoroughly as the Monocular section was. 

4. A review paper should also help to introduce new concepts that come from the current trending techniques. The reviewer feels it would be appropriate to generally explain at the beginning what appearance, motion, depth and expansion-based methods are. In all of those cases, it is described what is used for those methods. There is a need of deducing what they mean as the subsection is read.

4. The reviewer sees incompleteness of the paper having many formatting errors. It must be revised very carefully, but such revision is encouraged for future publication.

Several errors for example.

- Line 142 and 143: what is the difference between “ground features” and “ground data”? The terminology used there is a bit confusing.

- Line 196, there is a missing space in “2021are”

- In table 3, the reference Huh et al., 2015 does not have parenthesis.

- It is suggested to add the reference number in the tables or use the Author Year format in the whole document. It would help the reader quickly search for the reference described in the tables.

- In line 343, what do the authors mean by “environment becomes complex”?

Author Response

Reviewer #4:

The manuscript gathers context about different image-based obstacle detection methods. The paper first starts with the division of two categories: Stereo and Monocular. Then, elaborates it by describing the four techniques (Appearance-based, Motion-based, depth-based, and expansion-based) as part of the monocular category. Later, stereo-based approaches are divided into Inverse Perspective Mapping and Disparity histogram-based methods. The author then moves on to the abilities required from an obstacle detection algorithm such as narrow and small obstacle detection, moving obstacle detection, obstacle detection in all directions and fast/real-time obstacle detection. Several articles were cited, annotating their usage, principles, strengths and the criteria they met. Furthermore, the authors elaborated more on the four categories that comprehend the monocular method and continued to talk about the stereo-based method for terrestrial and aerial robots. The paper concludes by addressing the fact that Monocular-based obstacle detection approaches use only one camera. They are simple computations and are fast. It is noted that it was implemented on aerial and terrestrial robot navigations. Thus, it is mentioned that for Stereo-based methods there is a need for using of powerful Graphics Processing Units. Finally, it is stated that future methods will focus on detecting narrow and small obstacles and that deep learning techniques have helped reduce the computation time, for both monocular as well as stereo-based obstacle detection techniques.

Unfortunately, the reviewer feels the paper has not reached the level of review paper publication, the organization is poor, and not diverse enough;

1. The beginning of the paper introduces the Stereo and Monocular methods and later moved on to illustrate the four techniques used by Monocular-based methods, following the tree diagram in Figure1. However, when it is the turn of a stereo-based method, the paper focuses on the stereo-based obstacle detection methods for terrestrial and aerial robots (Instead of IPM-Based method and Disparity Based method).

Answer: Sorry. This was actually a typo. Actually, the grouping was fine (please refer to line 352 in the original paper), but the titles of the following subsections were wrongly named. The title of the sub-sections 4.1 and 4.2 should have been Disparity histogram-based … instead of Stereo-based…. In the new version, this mistake was resolved (lines 488-525 and lines 525-673). Thank you for noting. 

2. The stereo portion does not align with the structure of the paper that the author has followed in the monocular portion previously, therefore, can cause confusion to the reader.

Answer: Sorry. You are right. As mentioned above, this was a mistake in section naming and is resolved in the new version of our paper.

3. The reviewer believes this can appear as the stereo method section was not described as thoroughly as the Monocular section was. 

Answer: You are absolutely right. This, of course, was due to the fact that the majority of the obstacle detection techniques today are monocular. Nevertheless, we should have described the stereo methods in more depth. Thus, to address your comment, the Stereo section was extended with more details. Thanks to your comment, it is now more in line with the rest of the paper. The changes made can be found in lines 488-525, lines 555-557, lines 565-567, lines 571-572, and lines 580-583.

4. A review paper should also help to introduce new concepts that come from the current trending techniques.

Answer: Thank you for your comment. Yes, you are right. This comment was also raised by the second reviewer. To address this, several parts of the paper were revised to describe the current trends and express our personal perspectives on the techniques discussed in the paper. Here is a summary:

• At the end of each subsection, the technique is concluded, and our perspectives on solving some of the most important issues relating to the obstacle detection technique are given.
• Subsections "3.5" and "4.3" were added to the manuscript (lines 447-482 and lines 673-691).
• In addition, a new section “5. Discussions and prospects” was added right before the Conclusions. In this section, the overall features of the techniques presented in our paper, are reviewed and analysed, and the future trends and prospects in obstacle detection are mentioned (lines 692-777).
• Also some parts of the Conclusions were revised to present more in-depth points of view and to suggest some lines of work to be pursued in future research (lines 808-811, lines 817-821, lines 837-842 and section 6.1, lines 843-872).

5. The reviewer feels it would be appropriate to generally explain at the beginning what appearance, motion, depth and expansion-based methods are. In all of those cases, it is described what is used for those methods. There is a need of deducing what they mean as the subsection is read.

Answer: Very good suggestion. We did this. At the beginning of each section, several lines were added that explains what features are used by the method and how it works. Thanks to your comment, the paper is now more coherent and easier to follow. For details, please refer to the following addresses in the revised paper:

• For Appearance-based methods: Section “3.1.” (lines154-161).
• For Motion-based methods: Section “3.2.” (lines 242-247).
• For Depth-based methods: Section “3.3.” (lines 287-300).
• For Expansion–based methods: Section “3.4.” (lines 382-388).
• For IPM -based methods: Section “4.1.” (lines 492-501).
• For Disparity-based methods: Section “4.2.” (lines 526-530).

6. The reviewer sees incompleteness of the paper having many formatting errors. It must be revised very carefully, but such revision is encouraged for future publication.

Sorry for the errors, and thanks for noting. In addition to correcting the errors mentioned below, we went through the whole paper and reviewed it critically to make sure no formatting errors remained. These can be viewed all over the revised paper (shown when the track changes mode is turned on). The paper is now in a better state in terms of its writing accuracy. Thank you.

Several errors for example:

- line142 and 143: what is the difference between “ground features” and “ground data”? The terminology used there is a bit confusing.

Answer: Sorry for the confusion. Actually, they refer to the same thing. To avoid misunderstanding, the term “ground features” was replaced by “ground data” (line164).

- line196, there is a missing space in “2021are”

Answer: Thank you. It was corrected (line 212).

- In table 3, the reference Huh et al., 2015 does not have parenthesis.

Answer: Sorry about that. This is now corrected. In accordance with the Remote Sensing Journal format, we replaced all Author Year formats in the tables with reference numbers in [].

- It is suggested to add the reference number in the tables or use the Author Year format in the whole document. It would help the reader quickly search for the reference described in the tables.

Answer: Thank you. Reference numbers were added to Table 3 and Table 9. Please refer to the previous comment.

- In line 343, what do the authors mean by “environment becomes complex”?

Answer: Actually we meant “surrounding objects” This was changed to “surrounding objects become complex” (lines 430-431).

Author Response File: Author Response.docx

Reviewer 5 Report

This paper reviewed image-based obstacle detection methods for the safe navigation of unmanned vehicles like car and UAV etc. The paper is overall well-written. The reviewer has only one suggestion for improvement of the paper as follows. Since many technologies were introduced in the paper, it would be more interesting for readers if the authors can categorize different technologies in terms of localization precision requirements, deployment costs etc. One of such new tables added to the text would be helpful for the selection of suitable technologies in deployment phase.

Author Response

Reviewer #5:

This paper reviewed image-based obstacle detection methods for the safe navigation of unmanned vehicles like car and UAV etc. The paper is overall well-written. The reviewer has only one suggestion for improvement of the paper as follows.

1. Since many technologies were introduced in the paper, it would be more interesting for readers if the authors can categorize different technologies in terms of localization precision requirements, deployment costs etc. One of such new tables added to the text would be helpful for the selection of suitable technologies in deployment phase.

Answer: Thank you for your comment. Unfortunately, not all papers included such information. Nevertheless, very good suggestion. To address your comment, we used all the info and experience we had to develop a new table that ranks/reviews different techniques based on their accuracy, speed and cost (Table 11, lines 766-777). Thank you.

Author Response File: Author Response.docx

Round 2

Reviewer 4 Report

The authors corrected all comments that I made in the first review. 

Below are minor errors found.

o   Line 492: “The PM-based methods weres primarly…”

A sentence at Line 558 looks like having a typo

o   Line 558: “Salhi and Amiri [112] pre a faster algorithm…”

Author Response

Thank you for your comments.  I have corrected them and additional issues I found. I highlight them in Yellow color. Thank you.