A Study on Building a “Real-Time Vehicle Accident and Road Obstacle Notification Model” Using AI CCTV

Round 1

Reviewer 1 Report

The authors set themselves a very important and very difficult task: real-time detection of traffic accidents and non-accidental roadblocks based on CCTV video streams provided by cameras placed on the routes. By notifying vehicle drivers within 5 km of traffic obstructions about detected accidents and roadblocks, the aim was to change their route and avoid congestion and secondary accidents due to traffic obstructions. To solve this task, a very complex system has been developed, the input of which is provided by CCTV cameras already found on the routes. The recognition of real accidents and road obstacles would be based on the neural network taught using deep learning with the images of accidents and obstacles that can be detected with individual cameras. The three-server system required for this was developed, trained, and tested with virtual cases. The tests revealed the difficulty of the task, which stems from the resolution of the cameras, the decreasing of the viewing angles, the weakness of the overall input video signals provided, and the small amount of training.

Their work is pioneering, the benefits are clear.

The article is well-structured, readable, but lacking background information is made difficult the easy understanding due to many abbreviations. It is recommended to insert a glossary or an interpretation extract to give the interpretation of acronyms and professional abbreviations. The language of the dissertation can be followed well, the use of more precise expressions in some sentences would make the dissertation smoother.

The content follows the usual structure of the articles, perhaps a more detailed literature review and more literature could strengthen the dissertation. A description of the developed hardware - software system is sufficient to present the work done. The examples presented illustrate the complexity and difficulties of the task undertaken. Table 2 deserves a more detailed description.

The accomplished work is large, the results are remarkable, the paper is valuable.

Some minor remarks:

I suggest removing [] brackets from Table 1.

In Figure 2. the rhomboid has no False output.

In line 147: ".. can be derived can be derived." repetition

In line 157: x, y, w, h, c   -    c meaning is not given

In (2): doubled + characters

In Figure 4. RetinalNet -> RetinaNet

In Table 2. the method-type "RetineNet ResNet-101-FPN" is repeated, the headline is not explained.

From the Author Contributions section the first two and the last two sentences can be deleted.

Author Response

Hello
Thank you for checking in detail despite your busy schedule.
All items checked have been corrected.

The edited content has been attached to the file.
thank you.

 

 

Reviewer 1 Modification Request	Modifications	Whether to modify
I suggest removing [] brackets from Table 1	We removed [] brackets from Table 1.	Modifications completed
In Figure 2. the rhomboid has no False output.	We added Flase output in Figure 2.	Modifications completed
In line 147: ".. can be derived can be derived." Repetition	In line 221, we removed repetition.	Modifications completed
In line 157: x, y, w, h, c   -    c meaning is not given	From line 207~ we added meaning c -> confidence.	Modifications completed
In (2): doubled + characters	We removed doubled + characters, in Equation (4).	Modifications completed
In Figure 4. RetinalNet -> RetinaNet	We fixed it.	Modifications completed
In Table 2. the method-type "RetineNet ResNet-101-FPN" is repeated, the headline is not explained.	We fixed "RetineNet ResNet-101-FPN"  to “ResNeXt-101-FPN" in Table 4.	Modifications completed
From the Author Contributions section the first two and the last two sentences can be deleted.	We deleted them.	Modifications completed

Author Response File: Author Response.docx

Reviewer 2 Report

The author proposes a model that detects abnormalities on the road based on images acquired through CCTV cameras deployed in the real-time environment and uses deep learning techniques to enhance accuracy and efficiency. An accompanying web-based application notifies the users if an abnormality is detected. The author has validated the model by deploying it in the real-life scenario and creating virtual accidents.

 The paper solves a very important and interesting problem by using deep learning techniques and effectiveness is described in detail through the experiments. However, there are few major issues which needs attention:

• One of the major issues of this paper is the writing style which is hard to follow. It suffers from many grammatical issues that need to be fixed. Many of the sentences are incomplete and inaccurate. I will recommend benefiting from the editing services. Some of the sentences are unnecessary long which could easily be broken down into small sentences. It has many typos. Please refer to the attached file for more details.
• Author need to provide slightly more background of the topic that could give the novice reader an appreciation of the use of this technology and successfully link the reader with the topic of study.
• The literature review is not up to the standard of a journal. There is a significant related work missing in the area and the author needs to refer to important literature.
• The author has use too many acronyms without explaining and some of them are not explained throughout the paper.
• The paper title contains the word AI which is never explained or described in the paper.
• On page 2, line 61, author claims that the proposed system prevents accident on the water but it was never discussed in the paper later on.
• Please explain that why YOLO is better than R-CNN and why former is preferred in the context of your work?
• What is the range of CCTV cameras? How many of them to install and how to select a location for installation?
• Table 1 is not properly formatted and some of the information provided is not accurate.
• Author has not provide much detail about the acquisition of the training data and how it was labelled. How long did it take to train the system? Similarly, details are missing for the test data.
• Placing the Figure 3 and loss function where they are described will be more helpful for the reader. I think the readability of the paper can be improved by proper labelling of the figures and relating it to the illustration in the text.
• It is not clear how the virtual accident condition was created in the experiment.
• Performance result for the experiments such as accuracy and time taken are not provided.
• The author has not described any limitations and design constraints. For example, the model performance during the nighttime. Does range and using wide angle has any impact on the performance.
• The novelty and significance of this work are questionable. There have been many works on road accidents detection, the presented problem does not seem to be unique to motivate new theory. At least, there should be a proper literature review.

Please refer to the pdf for minor issues.

Overall, I do believe there is value in this work but the magnitude of the contribution is not significant and also not well placed. The author needs to refer to state of art literature to improve the related work and highlight what is novel about their approach. This paper can be accepted if the issues highlighted are addressed properly.

Comments for author File: Comments.pdf

Author Response

 

Hello
Thank you for checking in detail despite your busy schedule.
All items checked have been corrected.

The edited content has been attached to the file.
thank you.

 

Reviewer2’s Modification Request	Modifications	Whether to modify
Author need to provide slightly more background of the topic that could give the novice reader an appreciation of the use of this technology and successfully link the reader with the topic of study.	In lines 53 to 65, we add a definition of the technology we use and the background that led us to choose this topic.	Modifications completed
The author has use too many acronyms without explaining and some of them are not explained throughout the paper.	We added descriptions for acronyms.	Modifications completed
The paper title contains the word AI which is never explained or described in the paper.	In line 29, we explained AI.	Modifications completed
On page 2, line 61, author claims that the proposed system prevents accident on the water but it was never discussed in the paper later on.	The content about that is a comparison of papers similar to this study. I modified it so that it can appear as intended, line 68 ~77.	Modifications completed
Please explain that why YOLO is better than R-CNN and why former is preferred in the context of your work?	We explained why YOLO is better than R-CNN, line 82~89 and line 195-197.	Modifications completed
What is the range of CCTV cameras? How many of them to install and how to select a location for installation?	We explained information about CCTV, line 384~391.	Modifications completed
Table 1 is not properly formatted and some of the information provided is not accurate.	We modified the format and contents of Table 1.	Modifications completed
Author has not provide much detail about the acquisition of the training data and how it was labelled. How long did it take to train the system? Similarly, details are missing for the test data.	Line 148 ~167, we added training data and test data information and we explained that we labelled focusing on car accident.	Modifications completed
Placing the Figure 3 and loss function where they are described will be more helpful for the reader. I think the readability of the paper can be improved by proper labelling of the figures and relating it to the illustration in the text.	We explained Figure 3 from line 142 to line 176. We added more information about loss function Equation (4) from line 224 to line 270.	Modifications completed
It is not clear how the virtual accident condition was created in the experiment.	We added the condition of first hypothetical experiment from line 384 to line 430. We added the condition of second hypothetical experiment from line 447 to line 484.  We mentioned the detailed conditions of the virtual experiment.	Modifications completed
Performance result for the experiments such as accuracy and time taken are not provided.	I don't express accuracy and time taken, but We added analysis of success and failure in the experiment, line 496 ~ 500.	Modifications completed
The author has not described any limitations and design constraints. For example, the model performance during the nighttime. Does range and using wide angle has any impact on the performance.	Experiment 3.1.1 established two of the actual CCTVs in Seoul and a route through them. The purpose of our experiment was to show that for all CCTVs managed by web applications, only users with the optimal route to the origin and destination entered are notified, not to be notified when an anomaly occurs. Experiment 3.1.2 Also, the purpose was to analyze CCTV in each of the three regions to show that abnormal phenomenon recognition is possible with actual CCTV images, and to analyze the reason if it is not recognized.	Modifications completed

 

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Author have significantly improved the manuscript