The Detection of Black Ice Accidents for Preventative Automated Vehicles Using Convolutional Neural Networks

Round 1

Reviewer 1 Report

Overall, an interesting study to detect black ice using CNN. It appears like the authors spent a good amount of time collecting and processing the dataset. The resulting accuracy of the model appears to be very high.

One immediate question I had was how this model be implemented in the real world. Is this going to go into the software that runs in the AV so that the vehicle can detect black ice and take preventative measures? Maybe this was obvious to the authors but it would help to make that clear. Having not extensive knowledge about AVs, I am also curious what preventative measures these vehicles can take. Do they just slow down or do something else?

I was also wondering (assuming the implementation plan above is accurate) if the the images should be obtained from AVs rather than general Google search. It seems to me that in Google, the images have been take specifically to show the conditions and so they will be in focus. The angle at which the image has been taken will also be different when taken from an AV I think.

Overall, it would be great if the authors spend some time describing how this model can be implemented either in the introduction or conclusion.

Here are some specific additional comments:

a) In the literature review, the authors describe a lot of studies involving image processing. There appear to be all kinds of details but I could not follow if there was a theme. I think the authors could organize it better and show relevance to the current study.

b) page 4, line 155: this sentence kind of assumes that CNN will work well for black ice detection. It says CNN are mode accurate and usable than any other black ice detection techniques. I believe the authors should substantiate this.

c) It is not clear how the number of images went from 2,230 (table 1) initially to 17,600 (table 5). Does data cropping and padding increase the number?

d) I feel like information in section 3.2 and Table 8 is duplicate to an extent. It should be good if authors described how these parameters were chosen instead.

Finally, I think the work done in this study has merit and would be useful in the future. However, I also believe that the authors can revise to describe and present the work better in light of my comments.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

This study employs CNN to develop an algorithm for classifying road surface conditions, including black ice, dry road, wet road, and snow road. The authors collected about two thousand image patches from Google Image Search to demonstrate the proposed algorithm. The results show that the proposed algorithm achieves good classification accuracy. The major comments for the improvements are listed as follows.

1, To my knowledge, there are lots of existing works focus on road surface condition detection based on artificial intelligence-based computer vision technologies (please see the following references as examples). CNN-based models are commonly employed in the existing studies. Comparing with the current study, the difference is the training data does not contain the category of black ice. Employing the original CNN model does not bring any novelty since the existing methods can easily achieve the goal of detecting black ice by adding a bunch of labeled images of the black ice category. Thus, I consider the current study is not suitable for publication due to the limited innovative contributions.

Pan, Guangyuan, et al. "Real-time Winter Road Surface Condition Monitoring Using an Improved Residual CNN." Canadian Journal of Civil Engineering ja (2020).

Carrillo, Juan, et al. "Design of Efficient Deep Learning models for Determining Road Surface Condition from Roadside Camera Images and Weather Data." arXiv preprint arXiv:2009.10282 (2020).

2, The introduction does not provide a clear picture of the research problem of the current study. The authors try to utilize autonomous vehicles as an eye-catcher to promote the paper for publications, however, describing irrelevant content does generate negative impacts for the readability which distract the reader from focusing on the main story. The authors should mainly focus on introducing the research idea from a theoretical perspective in the introduction section.

3, The literature review section presents lots of irrelevant works. The main objective of the current study is detecting black ice using image data. Thus, the literature review section should focus on the topic of black ice detection. The successful implementation of CNN techniques in traffic detection can not prove it should be useful for black ice detection.

4, Seen from sample data presented in the manuscript, the video sensor is very close to the ground. I am not sure if the data can be representatives from the implementation perspective. I recommend the authors test other data set to see how the performance vary from data sets to data sets.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Good paper with an important application to safety of automated vehicles. A couple of minor points:

• Please remove acronym (CNN) from your title
• Please extend the introduction, it is a bit short
• Please rename the "utilisation of CNN in transport" subsection to "deep learning applications to intelligent transportation" or something similar
• Please add citations to robotics literature which uses floor textures (cameras looking at the ground) to enable autonomy e.g. "A framework for infrastructure-free warehouse navigation" and "Early Bird: Loop Closures from Opposing Viewpoints for Perceptually-Aliased Indoor Environments" and "StreetMap - Mapping and Localization on Ground Planes using a Downward Facing Camera"
• Please remove Figure 7 (true/false positives) - it is basic knowledge and unnecessary

Author Response

Please see the attachment

Author Response File: Author Response.pdf