Marine Extended Target Tracking for Scanning Radar Data Using Correlation Filter and Bayes Filter Jointly

Round 1

Reviewer 1 Report (New Reviewer)

The authors have significantly improved the article. However, we continue to find the phrases "we" and "our". Fig. A1 mislabeled the coordinate axes. The x-axis is always directed north in navigation because directions such as bearing and course are always measured from north clockwise. After making the indicated corrections, the article is suitable for publication.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

The paper ”Extended Target Tracking for scanning radar data using correlation filter and bayes filger jointly" proposed a new hybrid method of a complex maximum likliehoods methods (e.g., correlation filter) partly supported by a minimum variance module such as EKF@Bayes filtermodule process average peak-to-correlation energy term. sacrificing its computation time, generally filter performance is better than comparing filters. I propose minor revision with requiring following revision.

 

1. Please move 3.1 and 3.2 to methodlogy section 2. 

particulalyr, 3.2 is explaining observation operator part. The observation operator part needs to be well described in M&M section.

 

2. Normally, this sort of filters require error covariance inflation or rejuvenization to avoid filter divergence, which requires parameter tuning.

Please explain how did you optimize the scale of such prior distribution information?

 

 

 in methodology, please briefly desribe the 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

Summed up, the work builds up on a joint filter, aka the JCBF, combining kinematic estimation and shape estimation based on an EKF and a KCF. There is a feedback from the fused state to the update step of the filters. The results, including comparisons, were clearly illustrated and explained. In general, the referee agrees that it is challenging to obtain constant and realiable features from a single sensor (as well as multiple) in the typical maritime environments. The applicability and real-time feasibility of the ETT (extended target tracking) method itself could thus be further explored within a wider scope in the future.

Some remarks are listed: 

Won´t it be more gramatically correct if "Joint Correlation and Bayes Filtering" is used? 

Lines 13 & 108: It is sound to assume that targets in automotive and maritime are originally of "rigid body" or "fixed dimension". What the authors refer to as varying shape are the target-born measurements due to the sensor resolutions, or reflection strength, which is nice to mention as well. 

Line 121: The hierarchy or structure of the paragraphs would be better if there would be no "Related work" since the other subsection starts directly after it.

Figure 1 + corresponding text: How many samples were required for the training? Does the number influence the response? 

Line 169: It is highly recommended to use Nearly Constant Velocity (NCV), see Bar-Shalom´s multiple works that set the basis of the recommendation. 

Line 187: So both filters are subject to double updates. What happens in the case no measurements arise from the target after consecutive scans?

Figure 2: How can the different axes of the response be explained, compared to the frame? 

Lines 191-195, Line 25 in Algorithm 1: How do the authors set the threshold that helps to identify disturbances?

Line 273: It is ambiguous here, whether the method developed was a single target tracker (ETT) or multiple target tracker (METT). Title and text don't match.

Figure 4 and general: Using ETT/METT in the title is a little ambiguous in the sense that in most applications where one wants to estimate the shape of a target, it is a lot more intuitive to have a Cartesian-equivalent representation. A vessel´s dimension is more understood in scalar values or metres. In this case, however, it could be a bit misleading as the true dimensions of the vessels are not directly estimated.

The referee is a little surprised that the works from Marcus Baum have not been covered in the introduction, being one of the major contributors in (M)ETT field based on both automotive and marine radars. 

Section 3.3.1: Could there be a diagram explaining the CLE overlaid on a true vessel for a clear illustration? The subsection mentions about calculating the centre's location of the vessels from radar reflections. In most cases with very large vessels (>100m in length) the estimated centre are very likely not aligned with the true centre of the vessel because some regions of the hulk could return more powerful reflections. This could also explain the "shape changing" phenomena. Therefore from what can be seen in Figures 6a-h, 7, and 9, it shall affect the metrics as well. So eventually, is it a question of how robust the tracker being used is at not losing the target, or on how accurate its estimates are despite of losing and re-capturing the target? 

Besides, what is the source (example, AIS) of the reference data?

Line 330: Why haven't the RM results been included in the plots? How was motion of the RM approach modelled? In practice if they were similar to the JCBF, the kinematic estimates should not differ that much that the target would be lost.

Finally, some proofreading will improve the overall text - especially for the second half of the article.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The authors have satisfactorily addressed all of my comments. The paper is in

much better shape now and I recommend it for publication in this form.

Author Response

Thanks for your comments and helpful suggestions, which are crucial 
to the improvement of our paper.

Reviewer 2 Report

Review of Revised version of

Marine Extended Target Tracking Using Correction Filter and Bayes Filter Jointly

By Jiaqi Liu, ZhenWang, Di Cheng, Weidong Chen and Chang Chen

 

The main content of the article remained unchanged. The article is about one new task in target tracking. Such a task is obtained when there are traditional observations of the range, azimuth, etc. and additionally there are visual (capturing the appearance features of the target) observations. There has been attention to this task for several years. Artificial intelligence technologies, in particular, neural networks, find a natural application in it. The authors note that there are no works on joint tracking using both types of observations.

 

The authors tried to combine solutions to two problems. The first task is the traditional task of filtering the stochastic system state by indirect observations. The second task is to classify targets by their images (scanning radar data). In addition, the authors have interesting material for experiments.

 

The authors made many changes during the revision of the article, but the main questions remained unanswered. Namely, the authors did not give the equations of motion-observations and did not justify the use of the Kalman filter.

 

More details:

 

1.     The authors write v_k target position, the state estimation v_k of the target's position, v_k is a fraction of x_k, x_k can be decomposed into two components as v_k and delta v_k. What is v_k and the more important question is: what is delta v_k?

 

2.     Where does the denominator in (15) disappear and what does the sign ∝ mean? Equality (18) is incorrect, because the expression of the joint density in (15) has a denominator. Why isn't he here?

 

3.     The authors write "Under the linear and Gaussian assumptions". Why aren't linear equations written for x_k, z_k instead of (1)-(2)? How are the estimates of the Kalman filter m_x and Sigma_x calculated? Note that the description of the algorithm says "Calculating by (20)", but without m_x and Sigma_x, formula (20) cannot be used. Write down the expressions for m_x and Sigma_x and (1)-(2) in the linear case.

 

4.     In the description of the experiment there is no main thing – the model of the observation system (1)-(2). Write down the equations of motion and observations.

 

5.     The description of the experimental data includes measurements of range and azimuth. These are non-linear measurements. How was the Kalman filter applied?

 

Therefore, the article continues to require substantial revision.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Second Review of Revised version of

Marine Extended Target Tracking Using Correction Filter and Bayes Filter Jointly

By Jiaqi Liu, ZhenWang, Di Cheng, Weidong Chen and Chang Chen

 

The authors have spent quite a lot of time revising this second version of the article. Unfortunately, the changes made not only did not make the article better, but led to the opposite. Inaccuracies have not been corrected, and more absurd errors have appeared in the text. The comments made earlier about the need to describe the model and the disappearance of the denominator in (15) were taken into account by the authors only with respect to the equations of motion. The revised version contains equations of motion with constant velocity, and instead of an incomprehensible reference to the linear Kalman filter, an extended Kalman filter appeared. At the same time, the authors did not bring the model to the end and did not give a formal description of the appearance measurement a_k variable. Nothing is known about this variable. As a result, the following gross errors remain in the work:

 

1. The denominator in (15) and above is a PDF of the vector A_k. Nothing is known about this vector. For example, we can assume that the components of this vector are discrete random variables and then the specified PDF simply does not exist.

2. In the previous version there was "Under the linear and Gaussian assumptions" to explain the relation (23). The authors abandoned the linear model, but left (23). The new explanation looks like this: "Under the Gaussian assumptions". In the nonlinear case, this equality is never fulfilled.

3. In the same way, equality (25) is never fulfilled. The replacement of distributions in Bayesian equalities with Gaussian densities is absolutely not justified in any way.

4. Appendix A the authors added in vain. Firstly, these ratios are well-known. Secondly, these descriptions are incorrect. Namely, the phrases "Calculating the covariance matrix of the prediction error" and "Calculating the covariance matrix of the estimation error" are incorrect.

5. There is a lot of "water" in the text. For example, the reasoning around (24). However, the fundamental issues are not clarified. First of all, the phrase after assumption (25) requires detailed explanation, i.e., justification of the relations for calculating m_2, \Sigma_2.

 

Therefore, the article continues to require substantial revision. In addition, the authors obviously need the help of a specialist in nonlinear stochastic filtering.