Novel Approach to Automatic Traffic Sign Inventory Based on Mobile Mapping System Data and Deep Learning

Round 1

Reviewer 1 Report

In this work, a methodology for the automatic inventory of road traffic signs has been presented.The methodology consists of four main processes: traffic sign detection (TSD), recognition (TSR), 3D location and filtering. I am worried the contribution of this paper is limited. This paper did not presented novel method for classifying the images more robustly or accuractely. 

Line 34. “Communication and mobility of people and goods are a key element of modern societies”. “are a”?

 

2.Line 41. “focusing more on maintenance existing ones”, maintenance existing ones ?

 

Line 49~51, “Related to this goal, new concepts called, Digital Infrastructure and Intelligent Transport System (ITS) are being developed [6]. Which have arisen in parallel to new concepts for mobility, such as fully automated infrastructures, electric, connected and autonomous car, and resilient infrastructures.” “….Which…” ?
Line 54. “good service conditions is” ?

 

Line 55~Line 57. “Nowadays, there are different techniques and technologies to achieve this goal (monitored digital 56 roads), the most used considering their effectiveness and applicability are techniques based on, 57 satellite images, aerial solutions and Mobile Mapping Systems (MMS).” Please re-written,

 

Line 58, “The solutions based on satellite platforms [7] show as main problem their low image resolution,” hard to read.

 

Line 62, “There are an emerging” ? an ?

 

Line62-63, “However so far … yet” ?

 

Line 71 “that is allowing to this technology to develop quickly” ?

 

Line 75, “which are a very relevant part”, are a ?

 

Line 77. “suggested show” ?

 

Line 80~84. “In particular, this work addresses the following specific objectives: Uses MMS data (images and point clouds) to inventory TS automatically. Takes advantage of each type of data in terms of time, accuracy and precision. Applies Deep Learning techniques to images for TS recognition and classification. Uses point clouds only to geolocate and eliminate false detections.” Please re-written these sentences following the format.

 

Above is just part of the language usage problem, please check the paper carefully or revising these paper by a native speaker.

Author Response

Please find the responses in the attached document.

Author Response File: Author Response.docx

Reviewer 2 Report

Dear Authors, I have reviewed the paper entitled "Novel approach to automatic traffic sign inventory based on Mobile Mapping System data and Deep Learning ". Please, see few of my remarks, Kind regards.   The paper deals with a method for mapping traffic signs based on Mobile Mapping System.
In my opinion, the topic presented in the Abstract is interesting to implement, especially during the continuous development of measuring systems and road modeling. In my experience it follows that mobile systems are great for this purpose because of the global position characters need not be high (<0.05 cm). Lines 58-64. Despite the accurate observation of the authors regarding satellite systems and air, I think that some literature should be added related to research of linear objects, e.g. in satellite systems, what resolution is considered as too low or either showing an example - airborne systems are commercially used for high voltage lines inventory. In my opinion, the description should be extended.   Line 101. I think that the spatial resolution depends on the vehicle's velocity and the type of device. I think that it would be necessary to rewrite the sentence. I agree, that photos are needed to uniquely identify the mark.   The article lacks the content of analyzes showing different conditions in which the measurement was performed, which could simultaneously help engineers/scientists dealing with this problem in other cities operating on different mobile systems. 
That's why I have the following questions:   How does driving speed affect the final result?
Is making the processes dependent on the density of points better to use the pictures themselves ? Focusing on photos is not innovative.

Author Response

Please find the responses in the attached document.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear authors,

Thank you for submitting this nice case study. I have several comments/suggestions that I believe must be addressed before the final submission/acceptance of the article.

 

English language: It is not bad, but there is quite a lot of sentences which are difficult to read with some linguistic and grammatical errors or wrong word ordering.  Please, involve a native speaker or a professional for help, as it will significantly increase the readability and clearness of your article. Additionally, it is very uncommon to start the sentences with the references or numbers in such a form ( e.g. [25], 2.5%). I would recommend you avoiding such cases. Introduction and state of the art: The review of the literature seems comprehensive and the motivation is quite sound. However, it is far far too long. To increase the readability of your article I would suggest reducing this part for at least 50%. Lines 113-132 (point cloud based TS detection recognition and positioning) can be written in 3 sentences. Lines 148-176 can be completely deleted and the article would not lose on its clarity. Your aims at lines 80-85 should be only 1 aim: fast TS detection, recognition and classification with accurate localization. Point. Figure 2. Unnecessary, does not bring clarity. Mathematical expressions, lines 255– 264, could be written clearer. From what I understood for each photograph you need to make manual camera calibration by clicking 4 points in the image. If this is correct the whole workflow is not automatic, faster and better/comparable to the state of the art and the main aim of the article is not fulfilled. Please clarify. Strange dollar signs appearing – Figure 4. Suggestion 1: Why do not implement some test to remove detected mirrors, when they are occurring so often? It should be easy and not too time consuming based on e.g. intensity info from point cloud. Suggestion 2: why don’t you estimate traffic sign position based on the pole and not plane and image center intersection, it would give much more precise positioning, without too much additional computation time Page 11 – you focus on showing less false positives, but you previously stated that false negatives are the bigger problem for real world application. You should compare that with other approaches Page 12 (383-385) -> you did not clearly analyse your positioning accuracy (how do you know you are better, how did you analyse this and what is your mean positional error ?) Suggestion 3: Incorporate in conclusion that your algorithm is 50% faster than similar (state-of-the-art) implementations.

 

 

Author Response

Please find the responses in the attached document.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear Editor

I have checked the revision of the paper, and the the author has revised the paper accroding to the comments. I agree the publication of this paper. 

 

Best Regards

 

 

Reviewer 3 Report

Thank you for addressing all of my points.