Vision-Based Highway Lane Extraction from UAV Imagery: A Deep Learning and Geometric Constraints Approach

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

I have carefully read your article entitled “Vision-Based Highway Lane Extraction from UAV Imagery: A Deep Learning and Geometric Constraints Approach”, which addresses a relevant topic and presents interesting results. Below I provide some comments and suggestions that could help improve the quality and clarity of the manuscript:

1. In the abstract, the three building blocks you list are well-established approaches in lane detection and segmentation. Presenting them as a “novel three-stage framework” is not convincing unless you provide a clear explanation of the technical advances or unique contributions of your approach.

2. While the bibliography is recent and includes appropriate journals and conferences, some references are not used properly. For example, references [22] and [23] are cited to support the claim that aerial analysis for autonomous driving is comparatively limited, although those works do not provide statistical metrics to substantiate that statement. Please review these connections carefully.

3. Although the introduction is structured, it does not provide a comprehensive review of similar works. Many recent studies address related challenges, but current advances are not sufficiently described.

4. Section 2 (Related Work) does not include any substantial discussion or bibliography, which weakens the positioning of your contribution within the state of the art.

5. The three points presented as innovative are not genuinely new; they have been applied in multiple previous works. However, the article as a whole may still present novelty through the specific integration or application context. A clearer contrast with existing methods is necessary.

6. In Section 4, there is no explanation of how hyperparameters were selected or whether other values were tested. This should be clarified to ensure reproducibility.

7. In the introduction, it is claimed that the proposed approach overcomes limitations such as lighting variations and image degradation. However, the data acquisition conditions are not clearly described, nor is it demonstrated whether these challenges are indeed addressed.

8. Given the large number of similar works, the discussion must be expanded. A comparison with other studies tackling the same problem is essential, not only in terms of performance metrics but also computational costs and differences in operating environments.

Overall, the manuscript addresses an important topic, but to strengthen its contribution, it requires a more thorough contextualization within related work, clearer methodological details, and stronger evidence supporting its claims of novelty.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The article addresses the challenges of lane detection in aerial imagery, such as occlusions, structural complexity, and adverse environmental conditions, by employing a multi-stage processing pipeline. The initial stage uses deep learning-based semantic segmentation to accurately extract roads and lane markings. This is followed by geometric processing that incorporates constrained polynomial fitting, allowing for outlier-resistant lane modeling. This comprehensive methodology proves particularly effective in managing real-world complexities, including solid and dashed marking patterns as well as segmentation inconsistencies. 

Below are my comments:

1. The 2x is not represented in Figure 2-1.
2. The specifications of the UAV and the camera should be provided.
3. The authors should clarify the relationship between Figure 2-1 and Figure 3-1.
4. The authors should explain the rationale of the spurious detections elimination when a lane marking simultaneously satisfies the following conditions regarding its vertical extent and curvature characteristics: (1) the normalized vertical span Ylen/iH falls below a predetermined threshold (as specified in Equation 3.2-4), and (2) the second-order polynomial coefficient c₂, which quantitatively characterizes the curvature of the fitted lane model, exceeds a critical value.
5. The authors should include references about data augmentation techniques, such as color jittering, random horizontal flipping, random cropping, and random resizing.
6. The justification for selecting an image resolution of 1080×720 should be provided, along with details regarding the batch size used during model training.
7. The computational setup used for training and evaluation, including hardware specifications, should be specified.
8. The number of images used for training and testing should be included.
9. The authors should compare their results with other methods to demonstrate the advantages of their model.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed all my comments. I consider the manuscript acceptable for publication.