Real-Time, Deep Learning Based Wrong Direction Detection

Round 1

Reviewer 1 Report

The paper presents a useful application in terms of remote road assistance. The adopted solutions are well known in the main literature. The authors should compare their results with other well know results from this field.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The topic corresponds to the scope of the journal.

The authors pose the problem of safe driving on the roads.

Line 123-125. Three main processes are not clear separated from each other. The authors can improve this text by providing delimiters, for example, 1) 2) or semicolumn.

Line 265: which was like the original paper results replace by was similar. Use more scientific language.

Recommended to revise English on the fact of the prepositions, many spelling shortcomings “the”, “a”.

Proposed system section: here the authors start to describe Figure 6, but then this process is interrupted by the introduction to updated Kalman Filter Equations, then seems they continue to describe blocks from Figure 6. The text needs to be re-formatted to be logical presentation of the proposed system. Authors need to decide they use term Equation or Formula.

Figure 9 is not necessary at all, doesn't make any sense.

Line 276: TCP, FTP

Refers to Figures and Tables should be done before they appear in the text, not after.

What do you mean flipped daytime? No justification that such method could be used as interchangeable for the wrong direction cars dataset was presented for both Case 2 and 3.

Table 4 is confusing: no explanation what does actual “no” or predicted “no”. What algorithm did you use for prediction?

Many photos that do not carry sense for the scientific paper.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper describes a system for detection of vehicles wrong direction. The system is based on deep learning techniques. The systems achieves 91,98% accuracy on average.

Overall, the paper is well written, but there many typos/wrong expression. Please, see some in the end of this review.

It is not clear why the authors didn't use available datsets so that results can be compared with previous work. Are these datasets available? Also, in the state of the art section, other works are described but the reader is left wih no ideia on the accuracy level obtained. This would be important in order to assess the results obtained in this work.

Section 2.2.2. Should be completely rewritten/extended. As it is, it is very difficult to understand; it seems just a draft version. The expression should be explained as well as the algorithm.

One question that arises is if we can take a model trained for a specific camera and use it on another one. Can the authors write something about this? In fact, ideally the paper should present results for different cameras (for example, in streets where also people pass). I understand that it may not be viable for this article, but the authors should write something about this.

The system is prepared to detect wrong "horizontal" direction. However, in Figure 1 there is a situation where it seems that vertical direction should also be considered. Can the authors write something about this?

Why aren't the results for the original night period results shown? If there is a good reason for this, it should be given. If not, these results should also be shown.

Typos/wrong expressions:

L56 - chapter -> section
L57 - described -> describe (you don't need to use the past tense here)
L73 - It is not clear the false positive detection in Fig. 2
L113 - 116: Please, rewrite this sentence. As it is, it is not completely correct.
Text in bullets should end with ; and the last bullet should end with .
L155 - "we chose to be fast" -> this expression should be improved.
L228 - "First was checking" -> improve this
L241 - "We annotated them using tool [15]" -> improve this
L251 - "We chose *time* various time..." -> delete
L262 - "mAP *of* 90,80%
L265 - "which was like the original paper results [16]" -> improve this

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The authors propose a video surveillance system which detects wrong way moving of vehicles on roads. The paper is well written, the scope and objectives are matched with the journal, the experimentation methodology is adequate, and the results are satisfactory.

Some observations can be made:

- Page 2, line 57: Although in Section 2 the description of previous algorithms contains figures with performance simulation results made by authors, the phrase “in this chapter, we described different techniques in computer vision to detect and track vehicles” may be replaced with “in this chapter, we described and implemented different techniques in computer vision to compare different detection and tracking algorithms for vehicles”;

- page 3: the paragraphs before and after figure 3 have different line spacing;

- Figure 5: mention the measuring unit for mAP. It is %? Also, the figures should be plotted using vertical bars or pies, like histograms, and not using lines between dots. The horizontal axis is discrete not continuous;

- line 119: “similar average precision” make confusion because the mAP is between 45% and 59%. To decrease the visual differences between mAPs, the vertical axis of figure 5 can be extended to 0 – 100% interval.

- line 121: why you don’t choose YOLOv3-320 which is faster than YOLOv3-416 (22ms instead of 29 ms)? Also, some comments about the mAP and Time performances function of the image resolution (SSD 321 vs 513, YOLO 320 vs 416 vs 608) should be made;

- figure 6: darker colors;

- page 7: description of Kalman filter contains a lot of typing mistakes!!!

-- Eq. 1 must be italic.

-- Eq. 1: u is not subscript, it is B*u. Also, y in Eq. 6, it is K*y;

-- Some variables must be Upper Case: p from Eq.2, s from Eq.4, k from Eq.5.

-- Eq. 5 can be written shortly: K= P * H^T * S^-1, where S is in previous equation.

-- Eq. 8 is without k.

-- the lines 181-183 are copied from https://en.wikipedia.org/wiki/Kalman_filter, together with Kk notation. On the wiki, the algorithm is described using time indexes k and k-1, but the authors simplified the equations writing without them. Also, the sentence “Proof of the formulae is found in the derivations section, where the formula valid for any Kk is also shown” should be deleted. In your paper, you don’t have the derivation section where you proofed the formulas!!!

- Before Eq.9 you should define the performance measure mAP = Detected cars / Num. of cars.

In table 3, the mAP from column 3 is 80/85 = 94.12 instead of 94.17.

- line 278: the video has a resolution of 807x646 pixels and the algorithm used in the training phase is YOLOv3-416 for a resolution of 416x416 pixels. Why you don’t use the YOLOv3-608? Comment this.

- Table 5: How are computed all performance measures? The values are in %? It is not clear the correlation with table 4.

- line 302: “cases achieved 91.98% of accuracy” -> “in the three cases, we have achieved an averaged accuracy of 91.98%”

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

As the targeted application is highly critical (involved human life) it would be useful to examine further the robustness of the method (e.g. different weather conditions, etc).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Some of the observation was eliminated: reduced number of figures, misspelling; but the most important ones related to the logical flow was ignored.

Again Figure 5 and section “Proposed system” do not have logical flow. Boxes in the Figure 5 are described not in the order. Calculate cost block is described before Hungarian algorithm block. Authors just changed the placement of the paragraphs in the section without modifying any text to make it more understandable. Process of describing the blocks is interrupted by the introduction to updated Kalman Filter Equations. It is better to present theoretical part when Kalman Filter is presented first time in the paper in the section 2.2

Table 4 is confusing with ‘no’ and ‘yes’, where ‘no’ means correct direction, and  ‘yes’ is wrong-way.  Should not it be the opposite? The same confusing with predicted ‘no’ and ‘yes’. There is no explanation what does predicted “no”, ‘yes’. For clarity use terms wrong direction and correct direction, and false-positive, false-negative terms for predicted parameters. What algorithm did you use for prediction?

Line 278: “inverted daytime and nighttime videos”. No justification that such method could be used as interchangeable to the wrong direction car was presented. It should be related works about that this method of inversion is effective for the training the data.

Change content, not just replace the paragraphs.

Reviewer 3 Report

Below, the authors find some simple issues to be solved in this version.

L71 - conditions, and as an example, -> conditions and, as an example,

Figure 3 is not referred in the main text.

L115 - As Figure 4 shows*,*

L123 - Racking -> Tracking?

L142 - The Kalman filter is *a* powerful tracking method

L155and156 - we focus on improving the time speed of detection, *robustness* in different lighting conditions, and *accuracy*.

L186 - pair. The process

Table 1, please correclty indent "Delete tracks"

L227 - the experiment -> remove "the";

L227 - At first checked -> The first consisted in checking...

L228 - and the second consisted in wrong-way

L241 - with *a* specific program

L246 - training so, in our case,

L252 - We chose time various time -> delete the first "time"

L265 - Time unit is missing

L278 - nighttime therefore, -> nighttime, therefore, OR nighttime. Therefore,

L279 - Figure 10 -> there is no Figure 10

Table 4 - Case 2: N = 114 -> they are 116

L293 - false-positive*s*

Author Response

Please see the attachment.

Author Response File: Author Response.pdf