A Real-Time UAV Target Detection Algorithm Based on Edge Computing

Round 1

Reviewer 1 Report

The present short study is interesting, the flow of the manuscript is good (though some parts should be re-arranged to become easier to follow and less hectic to read), and the set objectives are clear and well achieved. I add a few comments regarding the manuscript, while I would also strongly advice the authors to perform an extensive English language revision and improve the syntax of their manuscript, as it was the main shortcoming of their document.

Line 3: “conventional methods are difficult to prevent effectively”, this sentence makes no sense.

Line 13: The NMS abbreviation (Non-maximum Suppression) has not been defined earlier.

Line 13-15: This sentence needs further elaboration, as no sufficient information are given within the abstract.

Line 15: The abbreviations of the mentioned metrics (mAP and FPS) should be given.

Line 28: The term “hazards” does not seem appropriate in this context. The previous paragraph presented potential barriers and limitations related to UAV use cases, not risks.

Line 56 - 71: Since a dedicated literature section is following the Introduction, I would suggest to move this paragraph to Section 2 (Related Work), and ideally present the novelty and objectives of the present research (currently in Lines 72 - 93) afterwards.

Line 95 - 103: I believe this paragraph is more appropriate to be moved up in Section 1, due to its generic nature.

Line 113: “However, the method does not propose innovative improvements to the multi-scale structure in this paper”, if the improvements refer to the referenced paper, then it should be switched to “in that paper”, as it otherwise makes it appear as if you are comparing it with your manuscript.

Line 134: “But the accuracy of the model is a little low” I suggest rephrasing to something more formal.

Line 145: I find no benefit in presenting this Figure by “breaking” the paragraph that explains it in two divided parts. It would be easier for the reader to have a complete paragraph before the Figure is presented, so I would therefore suggest “merging” the two parts into a solid paragraph prior to the Figure.

Line 156: In Section 4, please check the entire paragraph again, as the purpose of various elements (i.e. the integrated modules, the activation function used etc) are described after their respective first mention.

Line 185: “the lightweight MobileNetV3 is a very excellent model” is not appropriate, it should be switched to something of neutral tone. I general, I advise you to avoid the use of words such as “very” and strong adjectives such as “excellent” in your manuscript.

Line 198: Use quotation marks to make it easier for the reader, as followed:

In Table 1, the term “Operator” represents the module used at this layer and the size of 198 convolution kernel.

Line 219: The word “scholar” is redundant.

Line 274: “and then the algorithm will 274 be trained and tested” Why is future tense used here?

Line 289: Add “The”…GPU is NVIDIA Quadro P4000…in the beginning of the sentence.

Line 294: Use plural (Hyperparameters) and adjust the sentence accordingly (eg Hyperarameters are a set of training-related parameters…..).

Table 3: For batch size, use the term “number of images per iteration”, instead of “training”. In “Weight decay” fix the typo, and remove “The” from the notes for consistency. Similarly, Momentum should be capitalized.

Section 5.3: Use plural (Evaluation Metrics), as not a single metric was used.

Section 5.3: Despite simple concept such as TP and FP have been described, other metrics such as IoU (which are arguably of greater importance for this paper) have been omitted.

Line 362: Delete “And” at the start of the sentence.

Table 6: Elaborate on the table headers.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

First of all, thank you for the paper and the nice work. Here are some comments to help you make the paper better. Please consider them.

It would be good to include references to works from research groups also developing drone detection algorithms, not just papers from the computer vision community, like: 
- C. Briese and L. Guenther, "Deep Learning with Semi-Synthetic Training Images for Detection of Non-Cooperative UAVs," 2019 International Conference on Unmanned Aircraft Systems (ICUAS), 2019, pp. 981-988, doi: 10.1109/ICUAS.2019.8797731.
- L. Mejias, A. McFadyen and J. J. Ford, "Sense and avoid technology developments at Queensland University of Technology," in IEEE Aerospace and Electronic Systems Magazine, vol. 31, no. 7, pp. 28-37, July 2016, doi: 10.1109/MAES.2016.150157.
- A. Carrio, J. Tordesillas, S. Vemprala, S. Saripalli, P. Campoy and J. P. How, "Onboard Detection and Localization of Drones Using Depth Maps," in IEEE Access, vol. 8, pp. 30480-30490, 2020, doi: 10.1109/ACCESS.2020.2971938.
- Opromolla, R.; Inchingolo, G.; Fasano, G. Airborne Visual Detection and Tracking of Cooperative UAVs Exploiting Deep Learning. Sensors 2019, 19, 4332. https://doi.org/10.3390/s19194332 

I argue that the usage of MoblieNetV3 is a novelty of your work as it has been used before. For example: Lisang Liu, Chengyang Ke, He Lin, Hui Xu, "Research on Pedestrian Detection Algorithm Based on MobileNet-YoLo", Computational Intelligence and Neuroscience, vol. 2022, Article ID 8924027, 12 pages, 2022. https://doi.org/10.1155/2022/8924027

The point is missing from the paper in my opinion is the connection in between the higher accuracy and the improvements one can gain through it in the real application. Does 0.1% make any difference, if you compare your results (Table 6.) with the pruned-YOLOv4 from Liu, H.; Fan, K.; Ouyang, Q.; Li, N. Real-Time Small Drones Detection Based on Pruned YOLOv4. Sensors 2021, 21, 3374. https://doi.org/10.3390/s21103374?

Also, why is important in this scenario to use edge computing? Please, write a short explanation to clarify this in the paper.

Please check the spelling of Multiscale-PANet. You can find various versions of it, like multi-scale PANet, Mutlscale-PANet, Mutilscale-PANet (Figure 2.). Please check spelling of other technical terms as well, like ReLU, MaxPool, etc.

Please correct Formula 1, using h-swish (with hyphen) instead of the current version using a minus sign.

The paragraph introducing the h-swish function (lines 212-215) makes little to no sense. Please, rewrite it. The h-swish is not the piece-wise linear version of sigmoid, but of swish.

Please correct Figure 4. - predict instead of prediect.

Please correct the sentence in line 243 (starting in line 242), "larger than an IoU" instead of "larger than IoU".

There is a logical hole in the paragraph from line 242 to line 249. Please reformulate the sentence in line 246 starting with therefore.

Also, rewrite the soft-merge algorithm in pseudo code style.

In line 355, please correct commen to common.

Please clarify why you used IoU=0.5 for mAP and how exactly APS, APM and APL calculated with regards to IoU=0.5:0.95.

Please explain more Figure 8. Why do you show the frame vs. fps, instead of showing a bar plot of the average fps and standard deviation for a long sequence? That way other aspects (like the effect of possible memory saturation or heat dissipation issues of the board) can be shown. In its current form that 10 frames tell very little.

Please make the console font size of Figure 9 bigger.

Please correct the References section removing unnecessary hyphens and including missing doi/arxiv data.

 

 

 

 

 

 

 

 

 

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The updated manuscript appears to have been significantly improved, and all comments from my initial review have been addressed in a very professional manner by the authors in their response. I suggest that the paper is ready for publication and can thus continue with the final revision phase by the editors.