Distance Assessment by Object Detection—For Visually Impaired Assistive Mechatronic System

Round 1

Reviewer 1 Report

This paper deals with an exciting topic. The article has been read carefully, and some minor issues have been highlighted in order to be considered by the author(s).

#1 What is the motivation of this paper?

#2 What is the contribution and novelty of this paper?

#3 What is the advantage of this survey paper?

#4 Which evaluation metrics did you used for comparison?

#5 It would be good if security domains for the deep neural network would be reflected in the related work such as BlindNet backdoor: Attack on deep neural network using blind watermark, Medicalguard: U-net model robust against adversarially perturbed images, Data Correction For Enhancing Classification Accuracy By Unknown Deep Neural Network Classifiers.

#6 In Figure 11, two screen result should be large to visualize the main point.

Author Response

REVIEW REPORT 1

This paper deals with an exciting topic. The article has been read carefully, and some minor issues have been highlighted in order to be considered by the author(s).

#1 What is the motivation of this paper?

Answer 1: Please, see lines 23 – 27, 57 – 60 and, especially, 61 – 64, 275 – 277.

 

#2 What is the contribution and novelty of this paper?

Answer 2: Please, see lines: 14 – 18, 61 - 63, 84 – 87, 96 – 100, 269 – 272.

 

#3 What is the advantage of this survey paper?

Answer 3: The main advantage of this research paper are evidenced in lines: 109 – 111, 275 – 277, 428 – 430 and, especially, 460 – 462, 523 – 528, 619 - 624.

 

#4 Which evaluation metrics did you used for comparison?

Answer 4: It is “mm”; please, see line 490 (Table 1) and line 505 (Table 2):

• from 500 up to 1500 mm for LIDAR and HD camera system – because of estimated distances for safety reasons for visual impaired;
• from 150 up to 300 mm for low cost camera – because of best performances of low cost cameras.

The evaluation was performed experimentally in the real environment on a set of 64 situations through the images captured by the cameras, by calculating the Distance Error (%), Average Frame (fps) and Class Accuracy (%).

 

#5 It would be good if security domains for the deep neural network would be reflected in the related work such as BlindNet backdoor: Attack on deep neural network using blind watermark, Medicalguard: U-net model robust against adversarially perturbed images, Data Correction For Enhancing Classification Accuracy By Unknown Deep Neural Network Classifiers.

Answer 5: Please, see lines 302 – 308 (hypothesis H5) and line 619 – 624.

The virtual prototype in Figure 1 is intended for use by visually impaired people. We consider that in the case of this device, directed by the user, the low speeds do not impose as necessary algorithms used for an autonomous vehicle. Therefore, the possibility of this device becoming the target of cyber-attacks such as those suggested by you is nil. The user has control of the device and is guided by it for daily trips, both indoors and outdoors. The device is helpful for the user, but it does not have the ability to make decisions for the user. Also, as mentioned in the article, the preponderance of the financial component was taken into account, the configured solution being a relative, low-cost and affordable one for ordinary visually impaired persons.

 

#6 In Figure 11, two screen result should be large to visualize the main point.

Answer 6: We have improved Figure 11.

 

The authors thank Reviewer 1 for all comments.

Reviewer 2 Report

Accept.

Author Response

REVIEW REPORT 2

Accept.

Answer: The authors thank Reviewer 2 for comments.

Reviewer 3 Report

The novelty of this manuscript is not clear.

Author Response

REVIEW REPORT 3

The novelty of this manuscript is not clear.

Answer: Please, see lines: 14 – 18, 61 - 63, 84 – 87, 96 – 100, 269 – 272.

The are two innovative aspects in this research. The first consists in distance assessment using low-cost cameras based on special signs detection (not used until now). The second stands in the development of a customized technical solution for perception subsystem using laser RPLIDAR sensor and photo camera.

Each of the above mentioned system was independently calibrated and the results evaluated.

The authors thank Reviewer 3 for all comments.

Reviewer 4 Report

This paper presents research results for distance assessment using object detection and recognition techniques.  The authors not only propose a new design but also provide practical prototype to support their design.  The reviewer is convinced the practicality of the system based on their experiences.  The paper is also well written.  The reviewer recommend that the article be published.

Author Response

REVIEW REPORT 4

This paper presents research results for distance assessment using object detection and recognition techniques.  The authors not only propose a new design but also provide practical prototype to support their design.  The reviewer is convinced the practicality of the system based on their experiences.  The paper is also well written.  The reviewer recommends that the article be published.

Answer: The authors thank Reviewer 4 for comments.

Round 2

Reviewer 1 Report

If possible, it would be good if security domains for the deep neural network would be reflected in the related work[1].

[1] BlindNet backdoor: Attack on deep neural network using blind watermark

Reviewer 3 Report

I don't have any other comments on it.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

1, the control strategy is not mentioned, for example,how the trajectory of the robot arm is controlled, how the trajectory is ensured to be optimal, and what control algorithm is used.

2, the innovation of this paper is not enough, the detection algorithm is quite common, the detection effect of YOLO algorithm and transformer algorithm in the recent target detection is better than the algorithm mentioned in the article.

3, multiple sensing fusion description is not clear, the article mentioned in the camera and radar is how to achieve spatial consistency?

4, how is the monitoring information applied to the control strategy?

5, the second part of the research is not coordinated, the logic is confusing, the second part is not highly relevant to the research, should review more literature related to target detection.

6, Lack of medical experiments, the medical application scenarios mentioned in this paper are not reflected in the experiments. It is impossible to verify the feasibility of the device in helping blind people to move the device.

7, The experimental scene is too simple, whether the visual detection works under various lighting premise, how to overcome the experimental conditions of lighting changes, and whether the accuracy of the detection is guaranteed?

8, There are grammatical and spelling errors in individual words.

 

Author Response

1, the control strategy is not mentioned, for example, how the trajectory of the robot arm is controlled, how the trajectory is ensured to be optimal, and what control algorithm is used.

R: We specifically mentioned the control strategy for the robotic arm is Directional Inverse Kinematics (see lines 34, 578)

In this paper, the focus is on object detection and recognition, not on control strategies (please, also notice line 2, 3)

 

2, the innovation of this paper is not enough, the detection algorithm is quite common, the detection effect of YOLO algorithm and transformer algorithm in the recent target detection is better than the algorithm mentioned in the article.

R: We mentioned YOLO algorithm and its implementation in our research (see line 333 – 337 and, especially, 338–341)

The innovation presented in this paper consists in two new customized methods for object detection and recognition, as well as for measurement the distance to the object (see line 29 – 33, 114 – 118, 128 – 130, 370, 427 – 429, 481 – 482, 510-513).

 

3, multiple sensing fusion description is not clear, the article mentioned in the camera and radar is how to achieve spatial consistency?

R: Spatial consistency is not the focus of this paper, as based on authors experience (with visually impaired persons) their capacity of receiving complex information is most of the cases, limited (see lines 372-372)

 

4, how is the monitoring information applied to the control strategy?

R: A Directional Inverse kinematics strategy of a serial robot is applied. No monitoring during motion as, based on authors experience (with visually impaired persons), monitoring all through the motion would be perceived as a stressing factor (see lines 377 – 378, 569 – 574).

 

 

5, the second part of the research is not coordinated, the logic is confusing, the second part is not highly relevant to the research, should review more literature related to target detection.

R: We have carefully checked the literature on target detection and, especially restructured the paper (see lines 226 – 232, 362 – 364, 375 – 377, 386, 473, 565, 588, 600, 620, 653).

 

6, Lack of medical experiments, the medical application scenarios mentioned in this paper are not reflected in the experiments. It is impossible to verify the feasibility of the device in helping blind people to move the device.

R: The object detection and recognition techniques were experimented in real environment. The obtained results are useful for these customized techniques implementation in a designed device for VIPs assistance.  (see lines 532 – 534, 566 – 567). The simulations with the robotic arm were performed in virtual environment (532-533, 565). The VIPs-Walker device is not a medical one, it is only an assistive device (see lines 35, 716).

 

7, The experimental scene is too simple, whether the visual detection works under various lighting premise, how to overcome the experimental conditions of lighting changes, and whether the accuracy of the detection is guaranteed?

R: We have explicitly added scene description (see line 379-385). Discussion on the obtained results on detection accuracy are also presented (see lines 593, 597). Please, also check the subchapters 5.1 and 5.2.

 

8, There are grammatical and spelling errors in individual words.

R: Authors checked and corrected English

Reviewer 2 Report

This paper deals with an exciting topic. The article has been read carefully, and some minor issues have been highlighted in order to be considered by the author(s). 

#1 What is the motivation of this paper? 

#2 What is the contribution and novelty of this paper? 

#3 What is the advantage of this survey paper? 

#4 Which evaluation metrics did you used for comparison? 

#5 It would be good if security domains would be reflected in the related work such as BlindNet backdoor: Attack on deep neural network using blind watermark,  Multi-Model Selective Backdoor Attack with Different Trigger Positions, Textual Backdoor Attack for the Text Classification System, and Defending Deep Neural Networks against Backdoor Attack by Using De-trigger Autoencoder.

Author Response

This paper deals with an exciting topic. The article has been read carefully, and some minor issues have been highlighted in order to be considered by the author(s).

#1 What is the motivation of this paper?

R: Authors clearly added the motivation of this paper (see lines 35-38, 48-51, 80-83, 88-96, 143-145, 719-721, 728-731).

 

#2 What is the contribution and novelty of this paper?

R: Authors have emphasized the contribution and novelty of this paper (see lines 28-35, 111-116, 125-126, 128-130, 357, 610-619, 736-737)

 

#3 What is the advantage of this survey paper?

R: This is a research paper. The concept and new customized techniques are to be implemented in new designed assistive device for visually impaired paper (see lines 715-718,762-764, 765-770)

 

#4 Which evaluation metrics did you used for comparison?

R: Metrics, but because the research is focused on two different techniques the units are different, meaning [mm] and [m] (see lines 539, 597).

 

#5 It would be good if security domains would be reflected in the related work such as BlindNet backdoor: Attack on deep neural network using blind watermark, Multi-Model Selective Backdoor Attack with Different Trigger Positions, Textual Backdoor Attack for the Text Classification System, and Defending Deep Neural Networks against Backdoor Attack by Using De-trigger Autoencoder.

R: Authors have considered and mentioned the backdoor issue (see lines: 782-786)

Reviewer 3 Report

Some comments should be concerned.

1. The contributions and novelty should be summarized, especially the differences with the related works.

2. The copyright of the figures should be considered, such as Fig. 9(b).

3. The section "Discussion" is redudant more or less.

Author Response

Some comments should be concerned.

1. The contributions and novelty should be summarized, especially the differences with the related works.

R: Authors have emphasized the contribution and novelty of this paper (see lines 28-35, 111-116, 125-126, 128-130, 357, 610-619, 736-737).

 

2. The copyright of the figures should be considered, such as Fig. 9(b).

R: There is no copyright needed. The Fig. 9(b) image represents authors’ mobile phone (see line 519). Authors have also checked the copyright box in MDPI submission system. It is also mentioned in the Copyright Submission Form submitted.

 

3. The section "Discussion" is redundant more or less.

R: Authors have restructured the Discussion section (see chapter 5, lines 609-713).

Reviewer 4 Report

An interesting device VIPs-Walker system for visually impaired people was designed by the authors.

 

However, I have the following concerns.

The writing style is not qualified for a scientific paper. The language is not concise, and the style is weird.

The novelty is far from sufficient.

The experiments toward the true visually impaired people are not finished.

Therefore, I have to refuse this manuscript.

 

Author Response

An interesting device VIPs-Walker system for visually impaired people was designed by the authors.

However, I have the following concerns.

 

The writing style is not qualified for a scientific paper. The language is not concise, and the style is weird.

R: The authors have improved the scientific paper (see revised version).

Their experience as researchers and, therefore, published papers is high.

Just mentioning “not qualified” and “weird” is NOT ENOUGH for reviewer’s report.

Authors respect his / her comments. but do not agree with them.

 

The novelty is far from sufficient.

R: Authors have emphasized the contribution and novelty of this paper (see lines 28-35, 111-116, 125-126, 128-130, 357, 610-619, 736-737).

What about authors’ comments like: “the review report is not relevant”?

 

The experiments toward the true visually impaired people are not finished.

R: The research activity is complex a one, takes times and, especially, from concept to validation there are many important and relevant phases.

This research paper presents concept, experiments and results for new customized techniques of object detection and recognition, as well as distance measurement.

The obtained results are used for modeling and simulation of robotic arm motion toward an identified and recognized object. For this, it is applied the control strategy based on

 Directional Inverse Kinematics, in virtual environment.

This paper shows the research and its results for the present state of the research.

Further research development is, for sure, to be done (please see lines 762 – 785).

 

 

Therefore, I have to refuse this manuscript.

R: Therefore, the authors, respect reviewer’s decision but, are completely astonished by the accuracy of his / her review report.

Round 2

Reviewer 1 Report

I would suggest the authors reconsider my previous comments and the two innovations mentioned by the authors are not solid as I said the customized detection method is not better than the state of art open source method. This paper may be reconsidered if author combined the control system together with the detection part and demonstrate a fully functional system

Reviewer 2 Report

 

The contribution of this paper is small.

The proposed method was applied to the field of classifying rifle datasets using the existing recognition model.

There is no new model or comparative analysis with existing studies such as various Yolo algorithm, and there are few specific details about the proposed model.

It is judged that the performance will be better when using the latest recognition model, Yolov5, etc., rather than the proposed method.

Reviewer 3 Report

I think it can be accepted in this form.