A Novel Traffic Flow Reduction Method Based on Incomplete Vehicle History Spatio-Temporal Trajectory Data

Round 1

Reviewer 1 Report

Author proposes trajectory restoration method using map data from three major cities. The results show improved error reduction performance of the proposed methods compared with the interpolation method in simulation stage. However, for the application of real autonomous vehicle tracking, the result that can guarantee the effectiveness of the proposed techniques in real time application should be included in the article.

Also, the traffic trajectory is very important data for safe driving. Therefore, the comparison between the error of retorted data using the proposed technique and real GPS data is required.

Another issue is that a real car not only moves sideways, but also moves in Z direction. What is the error due to in z-direction?    

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper deals with filling missing vehicle trajectory information via proposing a novel data-driven method. Numerical results show good performance with respect to two baseline methods.

• To start with, the scope of the presented work is rather vague and I cannot see it filling a precise gap in existing literature. There is a very large amount of research devoted to trajectory reconstruction, using various method, many of them machine-learning-based. Here the authors throw a new method, without however clarifying why there is need of such a new method and what are the weaknesses addressed by it.
• Algorithm 1 presented in Section 4.1 is claimed to have a complexity of O(n), because of only 1 loop present in the algorithm. However, two steps are not carefully explained: in line 5 there is a “search” command, which, depending on the data structure employed, implies a non-negligible computational complexity; in line 11, retrieving the maximum in a list (or queue) also has a non-negligible computational complexity.
• Same as for previous comment applies for Algorithm 2.
• Looking at Fig.5, it is clear that the method is designed for reconstructing trajectories in road networks with bifurcations and merges. However, the method(s) deals only with position and velocity, without clarifying how those are handled.
• The numerical results are compared with naïve methods, namely a KF-based approach and an interpolation approach. However, there are more advanced and well-performing methods as state-of-the-art that are neglected. Comparisons should be performed with such state-of-the-art methods.

Language throughout the paper should be revised to improve clarity.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The subject of the manuscript is interesting. After reading the text carefully, the following points were formulated. 

- line 15: "the" -> "The";

- lines 238 - 240 "For example, the higher the velocity value, the better the road network traffic condition, and the higher the velocity value, the worse the road network traffic condition." - error in the text: it should be written "the lower the speed, the worse the traffic conditions, ie congestion"; 

- algorith 2, 3: "poistion" - there is a typing error; correctly it is "position"; 

- How does the presented method compare with other similar methods? 

- what are the disadvantages of this method? 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

The title of the paper is wrong because

• the presented method analyzed the traffic FLOW and proposed a flow reduction method, not a traffic reduction method (under citation this if OK)
• speaking about vehicle trajectory, mentioning history and spatial-temporal features are evidence – they are not necessary for the title; it prolongs the title unnecessarily

Therefore, the suggested title is “A Novel Traffic Flow Reduction Method Based on Incomplete Vehicle Trajectory Data”.

The name of Institution 1 is misspelled: Technology should be in capital. Similarly, Institution 2: Software should be in capital.

Line 15: The sentence starts with a capital letter.

Line 19: Don’t start a sentence with “And”.

Line 30: “language of trajectory” sounds a bit strange

Line 36: Point of Interest is abbreviated as POI, in plural POIs

Line 41: unnecessary and wrong repetition of “algorithm to accelerate clustering”

Line 42: The sentence has no predicate.

Lines 46-47: The sentence has no predicate.

Line 56: “predication” should be replaced with “prediction”

Line 59: “recur” could be written completely: “recurrent”

Line 61: What is CA stand for? Please write it before being abbreviated

Line 76: Don’t start a sentence with “And”.

Line 85: Future tense is not necessary, so skip the word “will”

Line 89: Don’t start a sentence with “And”.

Lines 94-98: The word “In” starting all sentences is unnecessary; it is grammatically incorrect

Line 97: “is introduced” is duplication so that it can be skipped

Lines 98-99: The sentence is grammatically incorrect

 Line 107: The subtitle is wrong; my suggestion is “Context-based vehicle trajectory analysis”. Trajectory contains already information about history…

Line 130: The abbreviation LSTM must be resolved

Line 139: The title seems to be incorrect; my suggestion is “The method of missing trajectory point completion based on interpolation” The title is too complicated; maybe a substantial simplification is suggested.

Line 147: the word “traffic” is unnecessary

Line 162: What is ALPR abbreviation stand for?

Line 193: Convolutional neural networks belong to deep learning, but it is unnecessary to complicate the expression: please skip “deep learning” before convolutional neural network…

Line 196: I don’t understand “sliding track” as well as “consumption time of sliding”

Line 197: The sentence has no predicate, it can be fused with the prior sentence

Line 199: the abbreviation LSTM is wrong; please use it correctly: long short-term memory

Lines 209-210: same error with LSTM abbreviation

Line 224: The word “History” is unnecessary in the title

Line 229: Longitude is usually abbreviated as “lon”, not “lng”. This is also for later occurences! I do not agree with the independent storage of the coordinates and time.

Caption 1: after (a) and (b) please replace “Analyze” with “Analysis of”

Figure 2: The title of axis y has “position” and “velocity”. These are vectorial features, furthermore the latter can be derived from the position, so it is a redundant storage. Do the researchers really want this in managing so much trajectory data?!

Line 239: correctly “and the lower the velocity”

Figure 3: The label in the lowest cell (top layer, last row, first column) should be “p2” instead of “p3”

Equation (2): the correct notation of a norm has a double bar (single bar is for absolute value)

Line 256: why the Manhattan distance instead of Euclidean? A short explanation would be great

Equation (3): what is v0 stand for? why is it used instead of v’ ? The given formula assumes p_i <= p’ <= p_j, if t_i <= t’ <= t_j.

Line 273: why is minute used as a time unit? Is it enough or has a too small temporal resolution?

The paragraph has no message about the axes of the diagram: pos = const, vel = const, or time = const!

Line 276: Is hour adequate or h:m:s would be better for expressing time?

Lines 278-280: the sentence is unclear for me

Line 284: regarding “small and sparse” – if we analyze a lot of vehicles with their lot trajectories, the storage will be unbelievably fragmented, won’t be?

Line 290: Point(loc, vel, time) seems a suitable tuple for storing trajectory information (although velocity is a redundant component)

Equation (4): The enumeration is expressed by … and the average has a usual notation with overline, so the same is suggested for average speed.

Line 302: I miss some information about the size of TDC from the section.

Figure 4: Lon is suggested for abbreviating longitude (instead of lng)

In Definition 1 and 2 (Lines 328 and 334) the notation for points e.g. p_i(time) or p_i(position) is not informative. Both express temporal and spatial extent.

Line 367: If the velocity is required in the data model, then the sentence could be reformulated here as “trajectory points without measured velocity”

Line 378: Missing reference: “where v_m is…” There is no v_m in the prior sentence

Algorithm 2: Spelling errors in row 5 and 7 (position)

Line 390: M_v(position) is time of loss… – is it really time?? In the following line M_v(time) is time of loss… there is OK, but not at the first occurrence!

Line 393: v_i is miss point ID – is it really an identifier or rather speed?

Line 409-419: If you use Kalman-filtering, why don’t you apply it both for position AND speed estimation? An explanation would increase the understanding.

Line 420: w is called system/process noise, not deviation

Line 421: A is not a transfer matrix; its name is state transition matrix

Line 422: although matrix B takes part of influencing the system state change, the simple “converts inputs to states” is wrong

Line 424: v has a name of observation noise, not simply error

In the above equations system state vector x or observation vector z is not explained. A good reference book should be cited in the text!

Line 430: the explanation “Lines 4-7” is the correct one, instead of “Lines 4-6”

Line 431: the same: correctly “Lines 8-18”

Algorithm 3: row 6 is misunderstanding notation! If the described algorithm gives back a result with the return term, only the Kalman-filtering results will be given in this step. I suggest simply: Calculation of M by Kalman-filter

Table 2: The presented table can be extended with more informative columns, like number of observed points, number of logged tracks, number of road segments

Line 449: 2012.4-2013.4 – is it a time interval? Between April 2012 and April 2013?

Line 451: The Table reference doesn’t require a dot, e.g. Table. 2. This is a consequential error in the later text!

Line 452: Application Programming Interface is abbreviated by capitals: API

Line 460: Deleted points can be marked by more attention-grabbing notation, like red instead of gray. This is also for the next figures!

Table 3 caption: ED should be written: Euclidean, then the abbreviation MEE. The dimension [m] is to write in the caption, and then the table is clearer

Line 496: Don’t start a sentence with “And”.

Figure 8: It’s hard to read the axis texts

Table 4: I would write the dimension similarly in the caption: %. Further, the precision (number of decimals) should be the same, e.g. 6.4, while justifying the numbers to the right. There is a value of 70%; is it an outlier or is it normal? Do you have any explanation for this value?

Lines 522-523: I can’t understand the sentence

Line 530: why do you list these values in decreasing and not increasing order?

Line 534: “From Fig. 9 can see” is grammatically incorrect

Figure 9: the figures are hard to read when printed the paper

Lines 540-541: I do not understand the sentence “… on the information of front-rear trajectory points of missing trajectory points”

Line 569: The given coordinates are in the wrong format. Please specify more precisely the northing and easting (which is which). You can follow Google’s notation e.g.: 39°53'56.9"N 116°28'05.0"E

Same as for Line 570

Lines 583-584: Backtracking is a reserved algorithm, so your name is a bit misleading!

Lines 588-592: There is no predicate in the sentence!

Line 600: the term “factor of transposition position information” is unclear to me

Lines 606 and 609-610 seem a bit contradicting. Who financed the research?

The general impression was that the documentation and the paper should be significantly increased. After major revision, the paper should be reviewed again.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

In the previous review round this reviewer identified several issues and provided a set of comments corresponding to a major review, namely it was expected to generate significant changes in the manuscript to address those comments. However, the actions by the authors have been minimal and the comments have not been address satisfactorily. See below my reaction to the Authors' response.

C1- It is still not clear what is the scientific gap addressed by this paper. In particular, the authors mention that "the existing literature has few methods for training and reconstructing trajectories with small sample sets". What are this methods? What are their weaknesses? Why do we need to develop a new method and publish its results? Identifying the proper research gap and declaring the contribution of a paper is the basis of a scientific publication.

C2- I understand the first part; however, the authors should then better clarify the data structures and use more proper jargon. That is, explaining how the datasets are constructed and replacing "Search" with a more appropriate term. On the other hand, I am still not convinced about the second part: the authors use the term "put", which implies the use of a map (dictionary). If the dictionary is ordered, then this command "put" does not have O(1) performance; if it is not ordered, then "get the p_i with maximum P_r value" does not have O(1) performance.

C3- If Figure 5 is "only used for illustration and there is no process of explaining HTB-p method", what is the reason for keeping it in the paper? At least, its presence should be justified and commented.

C4- This is again not convincing (and strongly related to comment C1). The authors repeat here that "there is less research on the training of small sample sets". Why not considering those works, even if there are few (even one is enough), as a benchmark for the proposed method?

C5- Language is still below standard for a scientific publication. I suggest a very thorough proofread by a native English writer.