Artificial Intelligence Based Object Detection and Tracking for a Small Underwater Robot

Round 1

Reviewer 1 Report

Problems about format:

1. I think figure 6 just exceeds the boundary of the document. Same problem appears in the rest of them manuscript. 

2. figure 7 might have been at the center of the page. 

Problems about content:

1. Why the structure of UAV is divided into two parts which tethered with wire? A synthetic structure of UAV design is more popular and reasonable in underwater application. 

2. Can the author specified the superiority or novelty of the proposed method and CNN model compared with traditional target detection algorithm? 

3. I think that the author should list the compurational complexity of the proposed method, as the hardwares equipped on the AUV are always not that powerful. The computational loss is always one factor that should be taken into consideration. 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

The article focuses on applying artificial intelligence to a small underwater robot’s positioning tasks. Even though the undertaken topic is very interesting, the submitted paper has too many drawbacks to be published.

Major remarks

The article is written in an unintelligible way. It lacks consistency and clarity. There are plenty of language mistakes and editorial inaccuracy (e.g. lack of a part of the text – line 291, wrong references to the figures – Figure 17).

The authors paid attention to unimportant issues regarding the paper’s subject. The robotic system should be slightly described in the results as a platform for experiments. What is more, the description of IMU is pointless. Firstly, the authors presented the formulas, but they can be used in the presented form. Inertial navigation of underwater vehicles demands the combination of various sensors, such as FoG, accelerometers and magnetometers, which output signals are processed utilizing the Kalman filter. However, even sophisticated systems, such as VN-200, are capable of calculating position for only a few minutes. The authors presented an unnecessary description of the accelerometers’ principle of operation, but only in Line 483 the authors barely mentioned about Kalman filter. From the research point of view, the critical information is how accurate is the deployed system – how reliable the results are. Additionally, the usage of accelerometers for vertical velocity calculation seems incorrect. The pressure sensor should be used in this case since it does not accumulate errors in time. The depiction of the water quality sensor is also irrelevant.

The authors should point out the possible utility of the presented system. It is tough to imagine where it could be utilized. The presented research can be considered laboratory tests, but further experiments should include demonstrations in a relevant environment. From my point of view, this is impractical because the vehicle needs to drag the floating platform, which reduces the possible depth of operation. What is more, it is dangerous since the vehicle can accidentally increase its depth and sink the floating platform. The authors should present operational conditions for the proposed solution (range, depth, etc.).

The literature review is very superficial. What is more, the authors compare their solution with other methods but don’t mention them in the Introduction.

The objective of utilizing the stitching method is unclear. It seems irrelevant to object tracking and navigation. What is more, the assumption of using homography is wrong. The homography can be used when the feature points lay on the plane or the camera only rotate around its axis of projects. In the presented solution, the feature points don’t lay on the plane, and the vehicle can move in any direction. What is more, the applied SIFT method is very slow. Some research points out that the ORB method is more convenient for underwater applications.

Any details about developed software are not presented. It is not mentioned if the authors used any libraries or developed their code from scratch. For example, feature detection, matching and homography matrix calculation can be performed by deploying the OpenCV library. If this library was used, the authors should have informed about it.

The title suggests that the presented method performs target recognition. But, the proposed method only detects the defined object on the image.

The authors concluded that the devised solution is capable of real-time tracking and navigation. But, the authors tested their tracking algorithm only on the dataset, without real-time experiments with the vehicle. Consequently, they didn’t implement the algorithm depicted in Table 3 (the algorithm’s output is not a segmented image). The presented algorithm is very simple. It can’t facilitate the rotation of the vehicle – in that case, two propellers should be used in the opposite rotation. Considering an underwater vehicle as a nonlinear and nonstationary object, the researchers often deploy sophisticated controllers, such as fuzzy logic controllers or neural controllers. In this regard, the authors should at least consider implementing a PID controller, especially due to the fact that they use stepping motors, which are nonlinear.

The result part presented from Line 487 is unclear. I refer that it is related to navigation. But any details about the experiment are not given. The authors should explain this part in detail since it is the main subject of the paper. What is more, the authors state in this excerpt that the goal of avoiding obstacles was achieved even though they didn’t describe this goal in the previous sections. Underwater vehicles’ navigation and obstacle avoidance constitute a very complex problem. Consequently, the developed solution should be extensively described and explained.

Minor remarks

Line 16 – non-deep methods can lead to complete occlusion?

Line 19 – navigate to the target or track a target?

Line 30 – while operating in-depth, a long cable is necessary

Line 57 – nowadays?

Line 64-68 – irrelevant to the subject

Line 111 – more serious?

Line 116 – does the presented method solve the problem of insufficient water light source?

Line 152 – the propellers affect each other while working

Line 162 – which environmental variables

Line 250 – is image enhancement based on deep learning?

Line 274 (and more) – the comma is repeated

Line 331 – capital letters

Line 334 – I don’t think that this is the purpose of target tracking

Line 353 – it is difficult to cope with full occlusion

Line 382 – Figure 17?

Line 387 – Softmax is not a loss function

Line 433 – Figure 17?

Line 446 – Activates… ?

Line 453 – information should be included in the Introduction

Line 486 – why does not process data of magnetometer?

Line 492 – in which way?

Line 493 – how is the combination achieved?

Line 509 – how is recognition performed?

Line 514 – transposition?

Figure 17 – the enhancement step didn’t influence the number of corrected matchings

Line 537 – the submarine may be matched?

Line 552 – this task was not described in the paper (it is not trivial and demands sophisticated control algorithms)

Line 564 – if the FPS was equal to 180, there would not be a difference in the position of the vehicle between the presented frames (100 frames are less than 0.5 seconds). Additionally, I think that 10 FPS would be enough.

Figure 24 – the figure is unnecessary

Table 9 – day-time set?

Line 627 – the real-time performance and high-precision navigation were not proven.

Line 641 – considering underwater vehicles, their buoyancy is essential, not weight only.

Line 659 – does FPS equals to 35 not meet the requirement of real-time performance in the case of underwater vehicle control? Additionally, the authors should present hardware and software capable of performing detection so quickly.

Line 679 – provide enough current? The information is not related to the subject.

Suggestion

The authors should consider if they want to publish the research related to the developed platform or proposed tracking and navigation methods. They mixed both in the submitted paper, but none of them was described correctly.  

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The manuscript can be published in Processes after proofreading.

Author Response

Please see the attachment.

Author Response File: Author Response.docx