Speed Bump and Pothole Detection Using Deep Neural Network with Images Captured through ZED Camera

Round 1

Reviewer 1 Report

In this manuscript, the author proposed an automatic color image analysis method using deep neural networks for detecting potholes on roads using images captured with ZED cameras. The experimental results have shown that a deep neural network used to detect speed bumps and potholes can capture the specific situation of Sierra Street in Guanajuato, Mexico while achieving satisfactory high accuracy. Therefore, detecting and describing these anomalies can help reduce the risk of accidents and damage to vehicles. However, the reviewer has some comments.

1. It is recommended that the references be listed in citation order.

2. It is suggested to number the first section of the introduction starting with 1.

3. The review of relevant research on detecting potholes and bumps in the second paragraph of the first section is too confusing, and it is recommended to divide it into paragraphs.

4. Line 218, “…, it used 37 filters of 5 x 5 pixels with an accuracy of 92.06.” should be “92.06%”.

5. It is suggested to adjust the notes of Figure 10.

6. Line 242-244, “However, increasing the number of filters to 38 with a 3x3 filter size leads to a decrease in precision to 83.18%, as shown in Table 1”. Table 1 refers to the review of approaches to detect speed bumps and potholes. Does not align with the meaning expressed in this sentence. 

7. Line 263-264, “Our proposal achieves an accuracy of 98.13%, whereas that of Maeda et al reaches only 77% as shown in Table 2.”. I am not sure which one is Table 2.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This article detects speed bump and road potholes using deep neural network.

1. Introduction section must conclude with the contributions of the proposed work. The related work must be separated from the Introduction section and research gap must be clearly specified.

2. Authors have not provided the architecture diagram for the methodology. Image preprocessing, segmentation is not included in the proposed work. Authors need to concentrate on these aspects.

3. What is the significance of figure 3? What is the authors contribution in this architecture?

4. Results section is poor. Comparison of results obtained with existing work is missing.

5. Conclusion section must have a overall discussion of the work and the future enhancements.

This article proposes a deep neural network methodology to detect speed bump and road potholes. The article has several drawbacks and cannot be considered for the possible publication at this stage.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Summary

This manuscript proposes a method for color image analysis based on deep learning networks and investigates the feasibility of the method in detecting speed bumps and potholes. The filter size, filter quantity and their training and validation times in a convolutional neural network consisting of seven interconnected layers were analyzed by the index of accuracy. The authors claim that the deep convolutional neural network proposed by the research is both lightweight and highly accurate, and has the potential to enhance the safety of driving processes in autonomous or human driver-driven situations, providing an effective solution for the detection of speed bumps and potholes in roads.

Major issue:

1)        In Section 1.1 Convolutional Neural Network architecture, for the convolutional layer presented in line 167, it has 37 convolutional filters of size 33 and no convolutional stride. The authors should have explained in detail how the convolutional layer performs convolutional operations without a convolutional stride, and explained the construction process and distribution of these 37 convolutional filters. In addition, Figure 3 cannot completely depict the convolutional neural network architecture employed in this study, authors should attach a table containing the output shape of an image with the size (3766723) after passing through each layer of the model.

2)        In Section 1.2 Hyperparameters tuning, the authors present the importance of hyperparameter tuning in networks. However, the definition of hyperparameters and their role in the network are not described in detail. In addition, no mathematical equations are given in the description of the research methodology in chapter 1, and the authors should add some equations to make the research methodology more theoretically supported.

3)        In Section 2.1 Features Visualization of Convolutional Neural Network, the authors state that what the network learns in training is less clear. The authors do not adequately explain the training principles of convolutional neural networks. Convolutional neural networks act as a 'filter' when training image datasets, and the weight of label-related information in the images increases as the network progresses, and the authors should consult relevant papers for further explanation.

Minor issue:

1)        This manuscript contains three tables, the authors have not named the latter two, which leads to misquotations of the tables, for example, the mention of "only 77%" in line 265 should be derived from the third table appearing in the manuscript, contrary to "as shown in Table 2".

2)        The authors are inconsistent in their description of the number of layers of the convolutional neural network used in this manuscript. The abstract, Section 1.1 and line 263 of Section 3 both state that the CNN uses 7 layers, but line 278 of Section 4 shows that the CNN consists of only 6 layers, please review this carefully.

3)        In line 159 of Section 0, how many images are marked yellow, unmarked and not fully marked in the speed bump dataset, respectively? The authors should provide an exact number.

Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors addressed all of the questions clearly. The revised manuscript is appropriate for 

Applied Sciences.

Author Response

Thanks for your kind comments

Reviewer 2 Report

Authors have not incorporated the suggestions provided on including the architecture diagram. The architecture diagram has not been included. Figure 3 is not the architecture diagram.

Authors need to provide a architecture/flow diagram for the proposed methodology.

Rest of the comments have been addressed.

Author Response

Dear reviewer,

We have added a flow diagram for the proposed methodology.

thanks for your kind comments