Vehicle Counting in Video Sequences: An Incremental Subspace Learning Approach
, Jose Portillo-Portillo 1, Aldo Hernandez-Suarez 1, Jesus Olivares-Mercado 1, Gabriel Sanchez-Perez 1, Karina Toscano-Medina 1, Hector Perez-Meana 1, Ana Lucila Sandoval Orozco 2 and Luis Javier García Villalba 2,*Round 1
Reviewer 1 Report
In figure 1, for the key block functions, the results after applying the author method should be clearly demonstrated. Sufficient test and evidence should be provided to demonstrate and justify the proposal methods work. Both complicate and simple cases should be provided. The author should explain what is the difficulty to the provided cases when counting the number of cars and how to solve it.
Author Response
Reviewer 1:
l Reviewer 1 says: In figure 1, for the key block functions, the results after applying the author method should be clearly demonstrated.
Answer: Figure 1 only shows the general schema of the methodology. The results of each step are demonstrated in the following subsections e. g. Figure 2 shows the results of incremental PCA block; Figure 2 to 6 depicted the process from thresholding to binary hole filling; from Figure 7 to 9 the process of area of detection line to vehicle counting is shown. Also, the results demonstrate the performance of proposed method using a public domain video called highway from changedetection.net.
l Reviewer 1 says: Sufficient test and evidence should be provided to demonstrate and justify the proposal methods works.
Answer: We provided the performance of proposed algorithm operating in four different scenarios and we have specified the challenges in each one of them from line 206 to 216. Also, we provided their respective results in Tables 1-4.
1. Video No. 1 (Front view, periferico): 901 frames, recorded at 25 fps // 36 secs.
2. Video No. 2 (Front view, ESIME puente): 837 frames, recorded at 25 FPS // 34 secs.
3. Video No. 3 (Back view, ESIME puente): 1026 frames, recorded at 25 FPS // 41 secs.
4. Video No. 4 (Front view, Highways): 1700 frames, recorded at 25 fps // 68 secs.
l Reviewer 1 says: Both complicate and simple cases should be provided.
Answer: We do not provide such cases since we only focus on counting vehicles under specific conditions. We aim to solve the most common challenges such as small camera jitter, sudden illumination changes in the environment and the camera auto exposure time. Of course, there are special cases where this methodology cannot be applied. Comments about this has been added in lines 230 to 238. But your suggestions will be considered for future works.
l Reviewer 1 says: The authors should explain what is the difficulty to the provided cases when counting the number of cars and how to solve it.
Answer: We do express the difficulty of the provided cases (video sequences) in the Results section in lines 207 to 214. We Solved the traffic flow direction, the size of the vehicle, which implies the “aperture phenomena” in motion detection, but we still have some troubles with illumination adjusting and we are still working on it.
Reviewer 2 Report
Authors present a real time system to automatically count vehicles in video streaming.
They implement the incremental PCA algorithm to overcome limitation due to illumination change or camera jitter.
Section 2 is not sufficiently clear to understand alone, the incremental PCA. Formalism can be improved by mentioning in text all the variables used in the formula (i.e. forgetting factor or matrix B^ that should start from I_n instead of I_m) and in general to provide a general sense of the algorithm.
The proposed approach appears promising, nevertheless a more effective experimental setup could provide a more convincing overall evaluation.
First, a higher number of video sequences could be analysed and compared with other online or offline systems (to be used as groundtruth) to provide a consistent ROC curve.
The influence of varying T and f in the final results (in vehicles counting) is not clearly highlighted in the paper.
Author Response
Reviewer 2:
l Reviewer 2 says: Section 2 is not sufficiently clear to understand alone. The incremental PCA formalism can be improved by mentioning in text all the variables used in the formula (i.e. forgetting factor or matrix B that should start from I_n instead of I_m) and in general provide a general sense of the algorithm.
Answer: A more detailed explanation about the algorithm formulation and all the derivation can be found in the original papers [14] and [15] that are cited in the paper. However, attended your request additional details have been added to the formulation of IPCA. The general sense of the algorithm is illustrated in section 3.1, specifically the IPCA process and outcomes are shown in Figures 2 and 3.
l Reviewer 2 says: The proposed approach appears promising, nevertheless a more effective experimental setup could provide a more convincing overall evaluation.
Answer: We provided the performance of proposed algorithm operating in four different scenarios:
1. Video No. 1 (Front view, periferico): 901 frames, recorded at 25 fps // 36 secs
2. Video No. 2 (Front view, ESIME puente): 837 frames, recorded at 25 FPS // 34 secs
3. Video No. 3 (Back view, ESIME puente): 1026 frames, recorded at 25 FPS /// 41 secs
4. Video No. 4 (Front view, Highways): 1700 frames, recorded at 25 fps // 68 secs
And we have specified the challenges in each one of them from line 206 to 216. Also, we provided their respective results in Tables 1-4. Also, the final results demonstrate the performance of proposed method using a public domain video called highway from changedetection.net. However, we will take your suggestion for future works.
l Reviewer 2 says: A higher number of video sequences could be analyzed and compared with other inline or offline systems (to be used as groundtruth) to provide a consistent ROC curve.
Answer: Due to the time limitations it was not possible to perform the tests, but we are considering your recommendation for a future work.
l Reviewer 2 says: The influence of varying T and f in the final results (in vehicles counting) is not clearly highlighted in the paper.
Answer: The choose of a factor of two is merely experimental and is compensated by the average of the consecutive frames. Similar results can be achieved with 2.5*(sigma) or even 3*sigma. But we considered 2*(sigma) is a good starting point in our experiments. sigma is multiplied by a factor of two because, according to the literature, it contains 95% of all the information in a normal distribution of data. In our case, each frame I_proj behaves similarly to a normal distribution with mean 0 (0 value pixels indicate no motion has been detected). Approximately, the remaining 5% of data in each frame is the “motion” we are interested in. However, a more detailed explanation has been added in lines 144 to 151.
In the other hand, the effect of changing the forgetting factor f is shown in Figure 4. Due to that, choosing a value of f close to 1 leads to bad detection results in our application (counting vehicles) but may be beneficial in other use cases.
Reviewer 3 Report
The paper presents a method and algorithm for vehicle counting in video sequences from roads with multiple lines. The methodology is based on incremental subspace learning for moving detection from consecutive frames of video sequences.
The paper has major deficiencies like:
- Introduction must be developed (there is less than one page from 12). More recent references must be cited.
- The novelty of the paper must be highlighted. I did not see it.
- The described methodology is not new. Common and very known techniques are used
- The Algorithm 1 used old references. What is new? Also it is not fully explained (the notations, formulas).
- Row 97 – d: how much is it in the application?
- Iproj is a matrix or a vector? The same notation?
- Why T=2σ?
- Eq. (3) is not correct written: σ – number, X matrix or vector.
- Eq. (5) is not correct written
- Experimental results are not convincible: accuracy 100% at different speed.
- There is no comparison with other papers.
- There are too few references (especially new).
Generally, the presented solution is obviously not new.
The paper must be rewritten
Author Response
Reviewer 3:
l Reviewer 3 says: Introduction must be developed (there is less than one page from 12). More recent references must be cited.
Answer: Additional information has been added to the introduction section.
l Reviewer 3 says: The novelity of the paper must be highlighted. I did not see it.
Answer: Comments about the novelty of the proposed methodology has been added in table 6
l Reviewer 3 says: The described methodology is not new. Common and very known techniques are used.
Answer: Thank you for your comment, nevertheless, the methods like IPCA, post-processing (filtering, thresholding, etc) by itself are not new. However, the proposed methodology has not been applied in the area of vehicle counting. This is the main contribution of this paper.
l Reviewer 3 says: The algorithm 1 used old references. What is new? Also it is not fully explained.
Answer: The IPCA algorithm was not modified. However, it has not been used for develop a vehicle counting algorithm that is the goal of this research.
l Reviewer 3 says: Row 97 – d: how much is it in the application
Answer: d = 320x240 = 76,800. This is defined in the Results section (row 216). As I_width = 320, I_height = 240.
l Reviewer 3 says: Iproj is a matrix or a vector? The same notation?
Answer: I_proj is a matrix (image) obtained from the IPCA algorithm, but should be transform in a column vector (like all frames in the methodology I_i =[d x1]) lines 156 and 157
l Reviewer 3 says: Why T = 2*(sigma)?
Answer: sigma is multiplied by a factor of two because, according to the literature, it contains 95% of all the information in a normal distribution of data. In our case, each frame I_proj behaves similarly to a normal distribution with mean 0 (0 value pixels indicate no motion has been detected). Approximately, the remaining 5% of data in each frame is the “motion” we are interested in. However, a more detailed explanation has been added in lines 149 to 150
l Reviewer 3 says: Eq. (3) is not correct written: (sigma) – number, x matrix or vector.
Answer: The equation (3) has been corrected attending your observation.
l Reviewer 3 says: Eq. (5) is not correct written.
Answer: The equation has been revised and a more detailed explanation of the variables has been added.
l Reviewer 3 says: Experimental results are not convincible: accuracy 100% at different speed.
Answer: The accuracy of 100% in all lines is only achieved in Video 1 which is the video with better conditions, but in the other three cases even with the worst conditions the proposed method achieves precision rates over 90%.
l Reviewer 3 says: There is no comparison with other papers.
Answer: A Comparison between some previously published papers has been added in Tables 5 and 6.
l Reviewer 3 says: There are too few references (specially new).
Answer: New references has been added.
Round 2
Reviewer 1 Report
In figure 1, for the key block functions, the results after applying the author method should be clearly demonstrated. Sufficient test and evidence should be provided to demonstrate and justify the proposal methods work. Both complicate and simple cases should be provided. The author should explain what is the difficulty to the provided cases when counting the number of cars and how to solve it.
//---------------------------------------------------------------------------------------------------
No further pictures or cases provided in the revised version
Author Response
· Reviewer 1: In figure 1, for the key block functions, the results after applying the author method should be clearly demonstrated.
Answer: Thank you for your comment, the Figure 1 was redrawing and a detailed explanation of each block was include in lines 118-147
· Reviewer 1: Sufficient test and evidence should be provided to demonstrate and justify the proposal methods work.
Answer: Thank you for your comment, several experiment with different environments where included in the revised version and depicted in figure 5, and the explanation of these environments are in lines 173-177 these experiments shows the ability of proposed method to detect movements under different conditions.
· Reviewer 1: Both complicate and simple cases should be provided. The author should explain what is the difficulty to the provided cases when counting the number of cars and how to solve it.
Answer: Thank you for your comment, In lines 242-264 explain the issues related to motion detection problems in vehicle counting, and how it was solved.
Reviewer 2 Report
Dear authors, thanks for your answers.
I still have some minor concerns.
Please rephrase the sentence in 92-94.
Please refer to the algorithm steps in a more uniform way (101-107)
Please correct the formula in Algorithm 1 step 2. B^ starts from In+1
Author Response
· Reviewer 2: Please rephrase the sentence in 92-94.
Answer: Thank you for your comment, in the revised paper the sentence was modified lines 92-94
· Reviewer 2: Please refer to the algorithm steps in a more uniform way (101-107)
Answer: Thank you for your comment, in the revised paper the sentence was attended in lines 99-101
· Reviewer 2: Please correct the formula in Algorithm 1 step 2. B^ starts from In+1
Answer: Thank you for your comment, in the revised paper the formula in Algorithm 1 step 2 was corrected.
Reviewer 3 Report
Observations have been partially resolved
Author Response
Thank you
