Anti-Occlusion UAV Tracking Algorithm with a Low-Altitude Complex Background by Integrating Attention Mechanism

Round 1

Reviewer 1 Report

In the work, the authors address the problem of occlusion of a target in low-altitude airspace by using both the SeNet targeting technique and the LSTM trajectory prediction scheme for trajectories. More specifically, if the target is occluded, the LSTM flight trajectory prediction network is used to forecast the trajectory of the aerial vehicle.

The document is very well written and could be published in its present form. It can undoubtedly inspire many professionals. Nevertheless, I have some objections.

To be honest, it bothers me a bit that in the manuscript there are not at least some general, even approximate, relationships, or derivations, or even considerations that would somehow delimit the various factors or put them in relation to each other (I do not just mean narrative or machine learning level, which are almost perfectly handled). All the reasoning of the work takes place at the level of the standards level more or less, of knowing/not knowing of the databases in question. For instance, it would probably be worth analyzing visual resolution versus drone speed, and potentially these factors in relation to the LSTM's reaction times (I mean its computation speed and predictions, or power consumption as well). 

It might be important to answer more deeply and thoroughly how the ability to predict, for example, relates to the speed of movement and so on.  Another question is, for example, whether the curvature of the trajectory is important, since, as it seems to be the case with drones, there is obviously not much that can be said about trajectory in a general and universal way. 

From a mathematical point of view, the trajectory appears as a continuous, but considerably jagged line. Which suggests that a significant acceleration or even stopping of the drone can be expected. What if, by chance, the drone remained invisible behind the obstacle in the longer term from the perspective of applying its strategic viewpoint? Then how does the predictive capability decrease with such its strategy of concealing itself, and let's say also with some substantial obstacle ?  

Some minor suggestion :

1. Literature references are not always placed accurately or in sufficient detail. Readers may not necessarily be able to navigate through the references and obtain information about the primary sources. Sample manuscript text is provided below along with some suggestions for references that should be reviewed by the authors.    In order to detect UAV in low-altitude airspace as early and as far as possible using computer vision, it is often necessary to implement long-distance detection and tracking of  UAV [reference is required], which causes small imaging size and weak signal [reference suitable].

2. Please explain the parameter W_l, W_e, or possibly make a legend of all symbols which would be useful. 

Author Response

请参阅附件。

Author Response File: Author Response.docx

Reviewer 2 Report

The authors discuss the problem of an anti-occlusion UAV tracking algorithm. The paper is for interest for the journal readers. The language must be improved. Are sentences which can not be understood. The paper is well structured.  The novelty of the work must be described in comparison with the state of art. For example contribution 3 (LSTM trajectory prediction network) is widely used. Which is new in your proposal? Or which are the introduced advantages? Statements like "So far, all algorithms have not been able to solve the problem of target occlusion" must be proven. The problem is not solved indeed, but are several works with great results. The main issue of section 3 is reproducibility. The section is presented as a general lecture on the topic. Are not presented design details. A reader can not reproduce your results based on this description. The comparison with the existing methods is not fair. Please compare your results with algorithm tackling the occlusion problem.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear authors,

 Let me thank you for your extensive work, which I have the honour of reviewing here. The article has the character of a scientific study and illustrates the correctness of your focus. I must point out that some parts are confusing for the reader and as if some parts were only formally aimed at the UAV. This shows only a one-branch explanation of the occlusion problems in UAV management and monitoring. This problem is very necessary when controlling medium-heavy UAVs at low altitudes when the prediction of obstacles and loss of GPS signal is the sum of the dangers that a UAV operator can go through during telemetry control. It has achieved many results in this area. I still want to ask you to redo some parts of the chapters, such as the abstract, introduction, etc. to demonstrate that the problems you are solving are operatively integrated into the problems of UAV - drones.

I ask the authors not to use the possible achieved results in the abstract, but to show the advantage of their achieved results. What method did they use? And which is a prerequisite for their study.

Line 49 - 58: Conclusion in the chapter Introduction the authors set their goals and explain what will happen in each chapter. My view is that this is unnecessary to write about what the reader will learn in the article. But I miss the introductory picture (demonstrating) which some of their problems will be a summary picture of the problem as they describe in the introduction. Therefore, I ask the authors to improve the Introduction chapter.

Line 135 - The authors set another target for solving the problem. We already cited the previous goals in the introduction line 49. It is not appropriate for the authors to write new and new goals in each subchapter. The essence of the article is thus lost. I ask the authors for a correction.

Line 162 - X ‘- is in mathematical phraseology as a derivative, but this is not the case in this problem. We can mark differently it or it is the standard for your output: H × W × C. Or is it an expression of a figure?

Line 166 - the output is already written as X. Is it possible that this is a previous dynamic action and now the stabilization of the process?

Line 166 - mathematical operator given as convolution is mentioned above x y is by default interpreted as CircleTimes [x, y]. In a similar practice, the convolutional sign is used ....... This is at the discretion of the authors.

Line 172 - This is a clearer wording in relation (2).

Line 192 - We are talking about a set of points A, which should correspond to Figure 5. It is necessary to add the name of individual parts in the figure. And the finite part of the set A.

Line 210 - formula (6) is completed by the total sum where I miss from - to. In the character S..

Line 251 - entry and exit flight trajectory is shown in figure no. 7 drawn especially. Authors can improve this tag with a more accurate character. As they are the same.

Line 299 - Figure 9 is an example or result of an experiment? It starts with the word Some ...

Line 317 - I do not understand this subchapter. Experiment on Integrated UAV Dataset. In the introduction, the authors talk about the detection of UAVs in various conditions, including animals, etc. We must lose the reader in the last chapters, just like me.

Line 332 Conclusion. In the beginning, the authors set 3 goals and again another goal in subchapter line 49. In the end, the authors do not evaluate them in any way. I ask the authors to rework the whole conclusion so that it corresponds to the achieved results and set goals.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

I have got only one question, maybe not related strongly to the paper. How is the convolution (1) implemented in the numerical procedure? Which algorithm is  used?  I'm looking forward for the answer, not in the text. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors revised the manuscript in accordance with the recommendations. The only issue remains the novelty description. These descriptions still not reflect the real contributions.

Reviewer 3 Report

Dear authors,

thank you for your additional modifications, I have no further comments.