Estimation of Lane-Level Traffic Flow Using a Deep Learning Technique
Round 1
Reviewer 1 Report
In this paper, I recommend accepting with various experiments and contents. However, a few minor revisions are needed.
(1) It would be good to clearly indicate the contribution points in this journal.
As a result of simply substituting a fingerprint image in an existing program and analyzing it, it seems that the contribution point is somewhat insufficient.
(2) It would be good to highlight the differentiation by adding content to the research related to fingerprint recognition.
(3) It would be a good idea to add to the related research on security vulnerabilities related to fingerprint recognition.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The article presents a system for automatic counting cars traveling on a multi-lane motorway and determining vehicle speeds. These parameters are calculated for all lanes, which allows the assessment of traffic intensity independently on each lane. As input information, the system receives an image from a camera above the road. The authors used the well-known YOLO object detection system (using Convolutional Neural Network) and Deep SORT for object tracking. The combination of these systems, supplemented with logic checking the criteria for selecting objects for calculations, made it possible to determine the assumed parameters. The authors described the system tests on recordings made in a specific, indicated place (National Freeway No. 1, Taiwan). Probably, tests were not performed under real conditions, i.e. in a system with a camera providing a video stream in real time. It is known, however, that YOLO and Deep SORT can analyze and track moving objects in real time. I believe that the text should be supplemented with just this information: what is the efficiency of the entire system (with a connected camera), which resolution does it support, how many frames per second etc. The authors did not provide numerical data on the correctness of vehicle detection, but actually provided verbal information "...successfully identified all vehicles in the input video" (section 5: Discussion). Were really all vehicles correctly identified? Even if the system did not make any errors, the number of detected vehicles can be given. One can also test the operation with lower quality video stream, such as in rainy weather or at night. I suppose, in that cases the accuracy will go down.
Figures 4 and 6 present the concept of limiting the scope of searching for vehicles to determine specific traffic parameters. The drawings suggest access to a camera placed over the road and looking straight down. But the presented sample video is a perspective view of the highway, where the indicated straight lines separating the lanes and rectangular areas of interest cannot be drawn. Was any conversion from perspective to aligned view performed? If you have a curve view and no transforms, this method can be difficult to apply.
I suggest supplementing the text with the indicated information. Also, please checking the language correctness. I found various mistakes that make reading the text a bit difficult. For example: "system which can counting vehicles"; "vehicle is closed to the virtual line"; "newer versions has been developed"; "The main contribution of YOLO v4 are, it established"; "first it uses Kalman filter predicts trajectory"; "which can assigned"; "number of vehicles pass through each lane"; "velocity of each vehicles"; "to avoid backed up" etc.
Other remarks:
Figure 1, 2 - missing arrows (lines instead of arrows).
Line 166: contains only a dot finishing the sentence.
Formula 2: acceleration as delta(velocity)/delta(position)? Should be delta(velocity)/delta(time) - please check formulas.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
