Artificial Neural Networks and Deep Learning Techniques Applied to Radar Target Detection: A Review

Round 1

Reviewer 1 Report

In this paper, a comprehensive review of the deep learning techniques employed for target detection has been reported.

I consider this contribution well-written and clear. The reported literature concerning deep learning techniques applied to radar detection is quite complete, thus I thank the authors for their effort.

If I had to find some improvable features of the paper, I would suggest adding a short description of the main radar working principles for such a kind of application, concerning for example the frequency and phase analysis for range, speed, and displacement recognition.

This can be of interest for a general reader and can improve the paper understanding.  In the literature, you can find a lot of papers reporting the principles of FMCW and CW radars which probably are the most employed working mode for short range applications. I suggest the following two papers for FMCW and CW radars, respectively, because they report a very compact theory description which would be appropriate for this review.

10. Cardillo, C. Li and A. Caddemi, "Millimeter-Wave Radar Cane: A Blind People Aid With Moving Human Recognition Capabilities," in IEEE Journal of Electromagnetics, RF and Microwaves in Medicine and Biology, doi: 10.1109/JERM.2021.3117129.
11. Lohman, O. Boric-Lubecke, V. M. Lubecke, P. W. Ong and M. M. Sondhi, "A digital signal processor for Doppler radar sensing of vital signs," 2001 Conference Proceedings of the 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2001, pp. 3359-3362 vol.4, doi: 10.1109/IEMBS.2001.1019547.

You can also mention different applications of radar techniques where deep learning networks are or might be successfully applicated as the automatic detection of humans in cluttered environments, HVAC control systems, radar systems for structural health monitoring, hand gesture recognition, radars for unmanned aerial vehicles detection, radar self-motion cancellation, or monitoring of worker activities.

Author Response

Dear Editor or Reviewer,

Thank you for your letter and for the comments concerning our manuscript. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied your comments carefully and have made correction. The point-to-point responds to the reviewer’s comments are listed as following:  

1. If I had to find some improvable features of the paper, I would suggest adding a short description of the main radar working principles for such a kind of application, concerning for example the frequency and phase analysis for range, speed, and displacement recognition.

Response: Thank you for your valuable advice. According to your comment, we have adjusted and added some description of the main radar working principles for RTD in the subsection 2.1 and a brief information of theoretical derivation in subsection 4.2 and formula (3,4).

2. I suggest the following two papers for FMCW and CW radars, respectively, because they report a very compact theory description which would be appropriate for this review.

Response: Thank you very much to provide these papers for us. We have carefully read these two papers and been enlightened by the theory description of these two papers. In addition, we have added these two papers into our references.

3. You can also mention different applications of radar techniques where deep learning networks are or might be successfully applicated as the automatic detection of humans in cluttered environments, HVAC control systems, radar systems for structural health monitoring, hand gesture recognition, radars for unmanned aerial vehicles detection, radar self-motion cancellation, or monitoring of worker activities.

Response: Thank you very much for your advice. According to your suggestion, we have added a new part, including about 10 papers in the introduction section to specific emphasize the different applications of radar techniques in human activities, including hand gesture recognition, vital signs sensing, etc. Thank you very much to point out the issues of not comparing the proposed method with other models.

Special thanks to you for your valuable 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. We appreciate for your warm work earnestly, and hope that the correction will meet with approval. Once again, thank you very much for your comments and suggestion.

Best regards,

Wen Jiang and Ying Liu

Reviewer 2 Report

In this paper, the author made a survey on deep learning in RTD. I have the following comments. 

1. The survey should focus on the deep learning method in RTD, but the authors introduced some traditional methods in RTD.
2. RTD is a general topic and there are many types of Radar signals and images. which types do the authors focus on? 
3. What is the motivation of this survey?
4. In the reference list, many typical publications in TGRS, GRS, ISPRS are ignored. 
5. For the datasets in RTD, some typical datasets are ignored. Furthermore, the author should give the access link to these datasets. 
6. For the methods mentioned in the paper, please authors also give the link of code for these methods. 

Author Response

Dear Editor or Reviewer,

Thank you for your letter and for the comments concerning our manuscript. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied your comments carefully and have made correction. The point-to-point responds to the reviewer’s comments are listed as following:  

1. The survey should focus on the deep learning method in RTD, but the authors introduced some traditional methods in RTD.

Response: Thank you for your comment. We do focus on the deep learning method in RTD, but we also introduced some traditional methods as the related work, for one thing, the deficiencies and challenges of traditional methods were revealed, thus, driving researchers to investigate deep learning for RTD, which is also one of the motivation of this survey; for another, by describing the principle of traditional methods to explain why deep neural networks can be applied to RTD. We have added a brief description to explain the reason.

2. RTD is a general topic and there are many types of Radar signals and images. which types do the authors focus on? 

Response: We focus on both types of radar signals and images. In fact, in the process of literature research, we also found that radar signals and images are common forms of data for RTD. In our paper, we consider various data types, including radar echo signals, Range-Doppler spectrum, Pulse-Range images, SAR images, PPI images, Time-Frequency images, etc. We also discussed different deep learning architectures with different data forms in subsection 3.4, and introduced the preprocessing and construction methods of different data forms in subsection 4.2.

3. What is the motivation of this survey?

Response: Recently, applying deep learning to RTD is considered as a novel concept, yet by now, there is no paper which comprehensively summaries and introduces the application and its development status. In fact, our team is working on the deep learning-based methods for RTD, including building semi-physical experimental system for RTD and publishing radar echo dataset. Therefore, we did a lot of literature research to form this survey. We hope that this survey can provide a reference for future studies and applications of deep learning in RTD and related areas of radar signal processing. Thank you for your comment, we have added some information in this section to emphasize the motivation of this survey in 1. Introduction.

4. In the reference list, many typical publications in TGRS, GRS, ISPRS are ignored. 

Response: Thank you for your valuable advice. We have selected and reviewed the recent articles on online data-bases (e.g., IEEExplore, Open Science Elsevier, Scopus, Springer and Researchgate), which includes some typical publications in TGRS, GRS, and ISPRS. But according to your comments, we also have added and reviewed about 10 papers from these classic journals. Thank you for your reminding.

5. For the datasets in RTD, some typical datasets are ignored. Furthermore, the author should give the access link to these datasets. 

Response: We are very sorry for our negligence of not providing the access link to these datasets. We surveyed some datasets, and selected the two most relevant datasets to introduce. We have added the corresponding link to these datasets (if they were provided).

6. For the methods mentioned in the paper, please authors also give the link of code for these methods. 

Response: Thank you for your careful work, we rechecked these methods and added the link of code (if they were provided).

Special thanks to you for your valuable 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. We appreciate for your warm work earnestly, and hope that the correction will meet with approval. Once again, thank you very much for your comments and suggestion.

Best regards,

Wen Jiang and Ying Liu

Reviewer 3 Report

The manuscript presents a review on methods for Radar Target Detection (RTD). A lot of papers have been published on this topic in the last years. To avoid overlapping or even duplication of reviews, bias of the reported items, transparency, etc. in the last decade the principle of registration of reviews has been accepted – see:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3369816/,

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2802597/ .

The authors have not declared that their review is registered and it is a point of view of the Editorial board of Electronics if such review could be published in the journal.

Generally, the presented manuscript is well written, the language is fine and understandable. The following particular questions, remarks and recommendations are addressed to the content of the manuscript without considering the above mentioned item, which should be taken into account by the Editorial team.

Questions, remarks and recommendation:

• The title of the manuscript does not represent exactly its content, since the addressed methods are not only in the field of deep learning.
• A systematic literature review should follow a predefined protocol, which includes (i) definition of the databases for research, (ii) inclusion criteria such as keywords found in the titles, abstracts and keywords of the studies, as well as time interval in which the referred studies are published, (iii) exclusion criteria, such as language (e.g. only studies in English language are included in the review). The authors should provide the protocol that they have followed for the presented review – the strict criteria that guarantee that a significant study in the addressed field is not missed (intentionally or not).
• The authors should provide explanation of all abbreviations at the place of their first appearance – e.g. “MTI” on page 3, Pfa and PD on page 4, etc.
• The diagram in figure 2 needs additional explanation. What is presented in the “layers”? What is presented with the “connections” (the solid lines)? What is presented with the gray rectangles (with and without text)? This figure rather provokes question instead of providing answers.
• Considering that the presented manuscript is a Review, subsection 4.2 should present “Data Processing and Construction” applied in the referred studies. The studies should be mentioned at the place where the respective methodology is described, which is currently not done.
• What is the difference between “Ntotal_targers” and “Nde_tarcets” in subsection ‘4.3. Performance evaluation’? This point needs additional explanation.

Author Response

Dear Editor or Reviewer,

Thank you for your letter and for the comments concerning our manuscript. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied your comments carefully and have made correction. The point-to-point responds to the reviewer’s comments are listed as following:  

1. The title of the manuscript does not represent exactly its content, since the addressed methods are not only in the field of deep learning.

Response: Thank you very much for your careful work. We also thought about this problem when we decided on the title. Undoubtedly, this survey mainly focuses on the deep learning method in RTD, but we also introduced some traditional methods as the related work, for one thing, the deficiencies and challenges of traditional methods were revealed, thus, driving researchers to investigate deep learning for RTD, which is also one of the motivation of this survey; for another, by describing the principle of traditional methods to explain why deep neural networks can be applied to RTD. In addition, we also introduced some artificial neural network-based methods which are the basic form of deep learning. Therefore, although the title dose not represent the content very exactly, we believe that all methods mentioned in this manuscript are related to deep learning methods.

2. A systematic literature review should follow a predefined protocol, which includes (i) definition of the databases for research, (ii) inclusion criteria such as keywords found in the titles, abstracts and keywords of the studies, as well as time interval in which the referred studies are published, (iii) exclusion criteria, such as language (e.g. only studies in English language are included in the review). The authors should provide the protocol that they have followed for the presented review – the strict criteria that guarantee that a significant study in the addressed field is not missed (intentionally or not).

Response: Thank you very much for your valuable advice on how to present a systematic literature review. In this work, we reviewed and selected the articles on online databases, namely IEEExplore, Open Science Elsevier, Scopus, Springer and Researchgate, most of them were published between 1990 and 2020. In addition, we searched them by keywords in the titles and keywords of the studies, and only studies in English language are included in the review. We will also follow these protocols in future review studies.

3. The authors should provide explanation of all abbreviations at the place of their first appearance – e.g. “MTI” on page 3, Pfa and PD on page 4, etc.

Response: We are very sorry for our negligence of not explaining the abbreviation of “moving target indication (MTI)”, “false alarm rate (Pfa)”, and “probability of detection (Pd)”. We also rechecked other parts of our paper in order to avoid other omissions. Thank you for your careful work again!

4. The diagram in figure 2 needs additional explanation. What is presented in the “layers”? What is presented with the “connections” (the solid lines)? What is presented with the gray rectangles (with and without text)? This figure rather provokes question instead of providing answers.

Response: Thank you very much to point out the unclear part in Figure 2. Figure 2 illustrates the major developments and application of deep learning method in RTD. We divided it up according to publication date and content (i.e. RTD in clutter, RTD in noise, and DNNs for various RTD with different data forms), which can correspond to the subsection 3.2, 3.3, 3.4 respectively. The rectangles represent the distribution of major workload in different years, different “layers” represent different content. According to the comment, we revised figure 2 and described it more detailed, which could make the input forms more intuitive.

5. Considering that the presented manuscript is a Review, subsection 4.2 should present “Data Processing and Construction” applied in the referred studies. The studies should be mentioned at the place where the respective methodology is described, which is currently not done.

Response: Thank you for your valuable advice. We described data processing and construction in the subsection 4.2 as a summary of data processing methods. These methods are applied in previous review literature in subsection 3.4.

6. What is the difference between “N_{total_targets}” and “N_{de_targets}” in subsection ‘4.3. Performance evaluation’? This point needs additional explanation.

Response: Thank you very much to point out the issues of not explaining these two parameters clearly. We also realized this would be a problem, so we redefined these two parameters as: N_{total_targets} denotes the total number of targets in the sample, while N_{de_targets} denotes the total number of non-targets in the sample. According to the reviewer’s suggestion, we have added additional explanation of these parameters in the subsection 4.3 Performance Evaluation.

Special thanks to you for your valuable 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. We appreciate for Editors/Reviewers’ warm work earnestly, and hope that the correction will meet with approval. Once again, thank you very much for your comments and suggestion.

Best regards,

Wen Jiang and Ying Liu

Round 2

Reviewer 2 Report

The author revised the paper and responded to reviewers' comments. But this paper can not be accepted. 

1. The title is too generous. The type of radar sensor and imaging manner should be specified.
2. The authors summarized the NN-based target detection method, but not show their experiment results on typical datasets.
3. The authors focus on two most typical target detection datasets but only provide the link of one dataset---IPX.  dataset. 
4. The author declared to provide the code link of those methods in the paper, but it is not found in the revised paper. 
5. I am not sure whether the authors applied those methods to one or two target detection dataset. If did, the author should compare the performance of the mentioned methods on the typical datasets. Otherwise the review is less meaningful. 

Author Response

Dear Reviewer,

Thank you for your comments concerning our manuscript again. We apologize for not being able to give you satisfactory answers to these questions. We have reconsidered your comments carefully again and the point-to-point responds to the comments are listed as following:    

1. The title is too generous. The type of radar sensor and imaging manner should be specified.

Response: Thank you for your comments. We also thought about this issue when we decided on the title. Undoubtedly, this survey mainly focuses on the deep learning-based methods in radar target detection, but these methods we discussed in this paper does not make a specific distinction according to the type of radar sensor or the imaging manner. In the latest edition, we reconsidered the content of this review and corrected the title to “Artificial Neural Networks and Deep Learning Techniques Applied to Radar Target Detection: A Review”. We hope the new title can present the content of the manuscript more exactly.

2. The authors summarized the NN-based target detection method, but not show their experiment results on typical datasets.

Response: Thank you for your comment. The ANN-based radar target detection models we discussed were evaluated on simulation data, and various input data were adopted in these methods, e.g., received signals, CFAR data, Pulse-Range maps, Pulse-Doppler maps, statistical parameters. We also have summarized the processing methods, data forms and data sources and presented them in Table 1 and Table 2. Therefore, a direct comparison of experiment results on typical datasets for all NN-based methods is difficult-to-operate and not feasible. The diversity of data forms and non-disclosure of datasets are the two issues that we focused on, which we have emphasized in this review paper.

3. The authors focus on two most typical target detection datasets but only provide the link of one dataset---IPIX. dataset.

Response: According to the investigation, we selected the two most typical target detection datasets to depict. In the first round of response, we have stated that if the dataset links were provided, we would provide it. So, we provided the links of IPIX database which is available. But for CSIR, unluckily, despite many attempts, we were unable to find an official or valid link of the dataset, including in the literature that we cited. For example, it is declared that http://www.csir.co.za/dpss/brochures.html is available, but actually, it is an invalid link now. From many of the literature we have reviewed, we found this dataset was mentioned many times, so we collected a lot of information about this dataset which was presented in the paper. Due to the specificity of this field, we have reason to believe that data may not be publicly available. Although the link of this dataset is not available now, we still confirm the CSIR dataset is one of the important datasets for RTD dataset.

4. The author declared to provide the code link of those methods in the paper, but it is not found in the revised paper. 

Response: We reviewed these research papers again and again, and we confirm that most of the authors didn’t provide code link. We also added the code link in the corresponding reference if the authors provided it, but only a few. In this review, we are more inclined to introduce the main opinions, algorithms and the research way of thinking of these methods, but for the code link, if the authors didn’t provide the link, we also couldn’t provide it. Data and source code are private. If readers really want to refer to the code, they can send email privately, we have provided accurate references to the literature.

5. I am not sure whether the authors applied those methods to one or two target detection dataset. If did, the author should compare the performance of the mentioned methods on the typical datasets. Otherwise, the review is less meaningful.

Response: We didn’t apply those methods to one or two radar target detection datasets in this review paper. A direct comparison for all methods is not possible due to the fact that they are evaluated on different datasets of different data forms. In fact, the diversity of data forms and non-disclosure of datasets are precisely the problems that the field is facing right now. we also discussed these issues at length in this review paper. Unlike common practice in the field of computing, we cannot compare the performance of these methods on the same dataset, nor do we think this is something that a review paper needs to do. We cannot fully agree that without comparison in the same dataset, this review is less meaningful.

Special thanks to you for your valuable comments.

Best regards,

Wen Jiang and Ying Liu

Reviewer 3 Report

Although the authors have considered part of the recommendations in my 1^st report, the most important points are not addressed. These are:  

• The title of the manuscript does not represent exactly its content, since the addressed methods are not only in the field of deep learning.
• A systematic literature review should follow a predefined protocol, which includes (i) definition of the databases for research, (ii) inclusion criteria such as keywords found in the titles, abstracts and keywords of the studies, as well as time interval in which the referred studies are published, (iii) exclusion criteria, such as language (e.g. only studies in English language are included in the review). The authors should provide the protocol that they have followed for the presented review – the strict criteria that guarantee that a significant study in the addressed field is not missed (intentionally or not).

Related to the 2^nd item, the authors have provided some explanation in the answer to the reviewer. However, this is not enough. The predefined protocol should be described in a separate section in the Review, so that all above mentioned items (i, ii, iii) become clearly stated.

I also confirm that the current title “Deep Learning in Radar Target Detection: A Review” does not present exactly the content of the manuscript.

In my opinion, the manuscript is not suitable for publication in its present form.

Author Response

Dear Reviewer,

Thank you for your comments concerning our manuscript again. We apologize for not being able to give you satisfactory answers to these two questions. We have reconsidered your comments carefully and have made correction. The point-to-point responds to the comments are listed as following:  

1. The title of the manuscript does not represent exactly its content, since the addressed methods are not only in the field of deep learning.

Response: Thank you very much for your comments again. In the previous version, in order to demonstrate the motivation of applying deep learning-based models to RTD, we introduced the traditional methods of RTD as the related work. In the section of methods review, we considered the artificial neural network-based structure as a basic form of deep learning and most of the methods mentioned in this manuscript are related to deep learning, so we collectively called them “Deep learning” methods. In the section of datasets summary, we summarized some data processing and dataset construction methods mentioned above. In fact, we also realized the title “Deep Learning in Radar Target Detection: A Review” was not the most appropriate one. So, in the latest edition, we have reconsidered it and corrected the title to “Artificial Neural Networks and Deep Learning Techniques Applied to Radar Target Detection: A Review”. We hope the new title can present the content of the manuscript more exactly.

2. A systematic literature review should follow a predefined protocol, which includes (i) definition of the databases for research, (ii) inclusion criteria such as keywords found in the titles, abstracts and keywords of the studies, as well as time interval in which the referred studies are published, (iii) exclusion criteria, such as language (e.g. only studies in English language are included in the review). The authors should provide the protocol that they have followed for the presented review – the strict criteria that guarantee that a significant study in the addressed field is not missed (intentionally or not).

Response: Thank you very much for your valuable advice on how to present a systematic literature review. We apologize that we may have misunderstood your comments earlier. So, in the latest version, we have added a new separate paragraph in the section 1. Introduction to special emphasize the predefined protocol and its importance, and all above mentioned items were included. We have added a new paragraph in the latest version as follows:

“This review focuses on articles in online databases, e.g., IEEExplore, Open Science Elsevier, Scopus, Springer and Researchgate. Recent articles (available by July, 2021) published in major journal of radar signal processing and major international conference on artificial intelligence at-tract more attention, namely, IEEE Signal Processing Magazine, IEEE Transactions on Antennas and Propagation, IEEE Transactions on Aerospace and Electronic Systems, IEEE Transactions on Geoscience and Remote Sensing, IEEE Geoscience and Remote Sensing Letters, IEEE Journal of Selected Topics in Signal Processing, IEEE Transactions on Signal Processing, ISPRS Journal of Photogrammetry and Remote Sensing, IEEE Radar Conference, International Conference on Radar, IEEE Conference on Computer Vision and Pattern Recognition, IEEE International Conference on Computer Vision. In addition, a number of research papers from other sources are related to this topic and thus included in this review, and most of them were published between 2000 and 2020. These papers were selected by keywords in the titles and keywords of the studies, and they represent a wide range of: (a) methods from theoretical derivation to application re-search, (b) detection background from noise to clutter, (c) applications from maritime target detection to human motion detection, (d) data forms from radar received echoes to PPI images, (e) comparative methods from neural networks to deep learning models. At last, only studies in English language are included in the review.”

Special thanks to you for your valuable comments.

Best regards,

Wen Jiang and Ying Liu

Round 3

Reviewer 2 Report

In this paper, the author made a revision, but they did not solve all of my concerns. As authors responded, they did not compare the results of the summarized method on a public dataset. It is a summary of the NN-methods on target detection, but not a scientific review. I have to reject this paper. 

Author Response

Dear Editor,

Thank you for your letter and for the three reviewers’ comments on our manuscript. In the past one month, the reviewers provided three rounds of comments. Most of those comments are valuable and very helpful in revising and improving our manuscript, as well as have the important guiding significance to our researches. We are very grateful for their hard work and we have studied comments carefully and have corrected most of them in the first two rounds of the comments. Luckily, so far, two reviewers (Reviewer #1 and Reviewer #3) have indicated that they are satisfied with our modifications and agreed to publish it. However, for the latest comment, I’m afraid that we can’t give a satisfactory answer to the Reviewer #2. So, we’ll explain this comment again to you.

In the recent review, the reviewer commented as follow：

“In this paper, the author made a revision, but they did not solve all of my concerns. As authors responded, they did not compare the results of the summarized method on a public dataset. It is a summary of the NN-methods on target detection, but not a scientific review.” 

In fact, we have responded to this question in the second round of comments, but apparently, the reviewer didn’t approve it. In this review paper, we investigated more than 100 papers, and more than 30 methods were discussed with emphasis, and most of the classical methods were summarized in Table 1 and Table 2. It is impossible for us to compare the results of the summarized method on the public dataset, because:

• 1. Most of the authors didn’t provide the code link. Some of these papers were published early without code for implementation, which was common in early days. In addition, the datasets they used were not uniform. For example, for those ANN-based methods we reviewed, simulation datasets were utilized separately to evaluate the models but without detailed parameters of the simulation datasets. In other words, even if we were able to reproduce the code and reconstruct the simulation dataset, there is no guarantee that we would get the desired experimental results as the authors did. Moreover, it would be very confusing if we get worse results than the authors.
• 2. According to our research, various forms of input data were adopted in these methods, e.g., received signals, CFAR data, Pulse-Range maps, Pulse-Doppler maps, statistical parameters, PPT images. Therefore, it is very difficult or even impractical to compare the results of the summarized methods on a public dataset. The diversity of data forms and non-disclosure of datasets are precisely the problems and challenges that the field of radar target detection is facing right now, which we have emphasized in this review paper. At least for now, we can’t break through the obstacles and reach a satisfactory result.
• 3. We have to state that this is a review paper, in which we are more inclined to summarize useful information, introduce the main opinions, algorithms and the research way of thinking of methods, explore opportunities and challenges, discuss development tendency, etc. We also followed a systematic approach to constructing this review paper. On the other hand, if we can reproduce all the methods, it would be very likely to improve on some methods, and then we would prefer to present the comparison results in another methodological paper rather than a review paper. We also believe that proper experimental results would make this paper more convincing, but obviously, the comparison results requested by the reviewer is not appropriate. We insist that even without the requested comparison results, this review paper is still of great scientific significance.

We tried our best to improve the manuscript and hope you can consider our explanation objectively. Regardless of the outcome, we appreciate for all reviewers and editors’ hard work earnestly.

Best regards,

Wen Jiang and Ying Liu

Reviewer 3 Report

The authors have considered all recommendations and the manuscript has become suitable for publication.

Author Response

Dear Reviewer,

We appreciate for your hard work earnestly. Thank you!

Best wishes,

Wen Jiang and Ying Liu