Research on HAR-Based Floor Positioning

Round 1

Reviewer 1 Report

The authors addressed all my previous concerns, answering my questions and making the necessary suggestions. The English problems were solved which makes the manuscript more easy to read.

I believe the paper is almost ready to be published after some minor revisions. I still have some small concerns and suggestions:

1 - In line 157, it should be "have achieved good performance”. Another grammar revisions should be made for minor adjustments.

2 - In line 191, " According" should be with small initial.

3 - Figure 3 is not the very common way to present results, as there are too many combinations. At least authors should denote that for example the number of features increase the lower the row. Also, the legend should be more descriptive (same in other Figures).

4 - Section 2.5 should present the detailed metrics for the best algorithm.

5 - In Table 2, PF needs to be explained at the time the table is presented.

Author Response

The authors addressed all my previous concerns, answering my questions and making the necessary suggestions. The English problems were solved which makes the manuscript more easy to read.

 

I believe the paper is almost ready to be published after some minor revisions. I still have some small concerns and suggestions:

 

 

Point 1: In line 157, it should be "have achieved good performance”. Another grammar revisions should be made for minor adjustments.

Response 1: Thank you for your valuable suggestion. We have revised it in the manuscript (Please see Page 4, Line 158). In addition, the English in this manuscript has been polished by a native speaker again with the revisions highlighted in yellow accordingly.

 

Point 2: In line 191, " According" should be with small initial.

Response 2: Thank you for your valuable suggestion. We have revised the manuscript. Please see Page 5, Line 192.

 

Point 3: Figure 3 is not the very common way to present results, as there are too many combinations. At least authors should denote that for example the number of features increase the lower the row. Also, the legend should be more descriptive (same in other Figures).

Response 3: Thank you for your valuable suggestion. The red area in Figure 3 shows the effect of different feature vector combinations on the classification results of each classification algorithm. We have added a gradient bar legend in the original figure to make the results more readable. Moreover, for a clearer analysis and comparison of the performance of each algorithm, the average, maximum and minimum classification accuracy results of each classification algorithm have been added. Finally, the font type and size in the figure were adjusted accordingly. Please see Page 7, Lines 245-246, Lines 256-258 and Page 8, Line 262.

 

Point 4: Section 2.5 should present the detailed metrics for the best algorithm.

Response 4: Thank you for your valuable suggestion. We have added the average, maximum and minimum classification accuracy results derived from different classification algorithms in Fig. 3 of the manuscript. These parts clearly show the classification performance of each algorithm, for the purpose of easy comparison and selection of the optimal algorithm. Meanwhile, the related words and sentences regarding this figure in section 2.5 were adjusted as shown in Page 7, Lines 237-241, Lines 245-247, Lines 256-258 and Page 8, Line 262.

 

Point 5: In Table 2, PF needs to be explained at the time the table is presented.

Response 5: Thank you for your valuable suggestion. We have revised the manuscript. Please see Page 8, Lines 276-277.

 

Author Response File: Author Response.docx

Reviewer 2 Report

The authors replied to my previous concerns. 

Author Response

Thank you for your comments concerning our manuscript.Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches.

Reviewer 3 Report

My comments have been been addressed

Author Response

Thank you for your comments concerning our manuscript.Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches.

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

This paper presents a solution to assist indoor positioning systems in the floor change detection, using both the barometer and the inertial data. Although the theme is quite relevant and is often missing in the literature, I believe the work presents some problems, specially regarding the coherence between the motivation and the developed work. At this stage, the manuscript is in no condition to be published, although it may be considered after major revision. My comments are the following:

• The presentation of the paper was not taken care of. While some figures do not have enough resolution, other have missing information, as labels explanation or values description. 
• The first phrase of the Introduction seems wrong. It seems that it was misspelled. English problems are very common throughout the manuscript, which must be corrected before submission.
• In the beginning of Section 2, the authors refer to "the" HAR algorithm when detailing the literature limitations, where the transmitted idea is that one general algorithm exists, which is not true. The goal is to achieve HAR, but many approaches may be taken. Again, the text is confuse.
• Many acronyms appear without a first definition of their meaning.
• In filtering, the “window of 9” refers to nine timesteps? Is there any overlapping? Why this value?
• The algorithm in Section 2.2 does not have legend and need proper explanation in the text. The authors use some thresholds, which are not even mentioned. They need to be explained.
• In line 176, the authors use 10.5~16, but the ~ does not mean "to", it should be properly written. Again, no explanation is given for these thresholds.
• Why did the authors decided to take a first threshold-based layer to make the distinction between lifts and other types? A general ML-based approach would not be more simple?
• In 2.4, the authors refer to the feature vector as the values of one single computation/feature. But the feature vector is the aggregate of all features.
• I do not understand how the authors obtained the 88 features if there were 28 features in 8 dimensions plus 3 more for the barometric data. 
• Table 1 is confuse. It should have one feature per line, maybe using two sided tables for space optimization.
• The authors say that the limitation of the literature are related to the use of Wi-Fi and barometer. But the authors also use the Wi-Fi for first floor initiation and the barometer for feature extraction. How do the authors surpass the limitations of the literature then?
• Also, the authors say that they will address the real-time detection, but in Section 3 they say that the floor positioning is done recursively.
• Futhermore, the authors say that the limitations of the literature are related with the need for maintenance and implementation, which harms scalability. But this system also requires an extensive detail of all floor plans.
• Table too also has parameters that are not explained.
• How many data points were used for the linear regression? Isn't the model overfitted to a single elevator and device?
• What is the meaning of "temp" table?
• What is the 'Activity results' od Figure 8?
• How do the authors assure that the system is not in general overfitted? If more than 3 consecutive misclassification happen, the whole system will fail.

Author Response

Response to Reviewer 1 Comments

This paper presents a solution to assist indoor positioning systems in the floor change detection, using both the barometer and the inertial data. Although the theme is quite relevant and is often missing in the literature, I believe the work presents some problems, specially regarding the coherence between the motivation and the developed work. At this stage, the manuscript is in no condition to be published, although it may be considered after major revision. My comments are the following:

Point 1: The presentation of the paper was not taken care of. While some figures do not have enough resolution, other have missing information, as labels explanation or values description.

 

Response 1: Thank you for your valuable suggestion. We have made major adjustments and improvements to the presentation part (Please see Page 1, Lines 14-17 and Lines 35-50; Page 2, Lines 61-70 and Lines 72-80; and Page 3, Lines 114-131.), improved the resolution of the graphics (Please see Pages 15-16, Figure 10-11.), and added content such as label descriptions (Please see Page 8, Figure 3; Page 12, Figure 7 and Page 14, Figure 9). At the request of another reviewer, we have added an experiment to compare our method with other HAR-based methods, which is reflected in the addition of a curve in Figure 10.

 

Point 2: The first phrase of the Introduction seems wrong. It seems that it was misspelled. English problems are very common throughout the manuscript, which must be corrected before submission.

 

Response 2: Thank you for your valuable suggestion. We have revised the manuscript through a professional editing service. Please see Page 1, Lines 34-35.

 

Point 3: In the beginning of Section 2, the authors refer to "the" HAR algorithm when detailing the literature limitations, where the transmitted idea is that one general algorithm exists, which is not true. The goal is to achieve HAR, but many approaches may be taken. Again, the text is confuse.

 

Response 3: Thank you for your valuable suggestion. We apologize that our description is not rigorous enough. Yes, our goal is to achieve HAR, and a variety of classification algorithms may be taken. And the relevant description in the manuscript has been changed. In some places, it was changed to "classification algorithm", and in some places it was changed to "HAR".  And we have revised the manuscript through a professional editing service. Please see Page 3, Lines 107, 139, 144, 150, etc.

 

Point 4: Many acronyms appear without a first definition of their meaning.

 

Response 4: Thank you for your valuable suggestion. We have supplemented and improved the manuscript. For example, please see Page 2, Lines 66, 68; Page 3, Lines 135, 145, 148, 150, etc.

 

Point 5: In filtering, the “window of 9” refers to nine timesteps? Is there any overlapping? Why this value?

 

Response 5: Thank you for your valuable suggestion. The window refers to the number of samples in the data. The average filter takes the average of the previous nine data points as the filtered data. The average filter can effectively remove the "noise" data and improve the efficiency of period division. The language description was also polished in the manuscript. The specific calculation process is reflected in Formula 2. Please see Page 4, Lines 173-178.

 

Point 6: The algorithm in Section 2.2 does not have legend and need proper explanation in the text. The authors use some thresholds, which are not even mentioned. They need to be explained.

 

Response 6: Thank you for your valuable suggestion. The overall algorithm content has been formatted, and corresponding instructions, notes and explanations have been added. The overall peak value determination logic is as follows: the filtered acceleration value effectively removes the false peak. If a sample value is greater than the three sample values before and after it, and at the same time is greater than a set minimum peak threshold, the sample is judged as a peak. Please see the contents of Algorithm 1 on Page 5 between Line 194 and Line 195.

 

Point 7: In line 176, the authors use 10.5~16, but the ~ does not mean "to", it should be properly written. Again, no explanation is given for these thresholds.

 

Response 7: Thank you for your valuable suggestion. We have modified the expression of the data interval, and added the graphic display of acceleration sensor data during pedestrian movement to assist our explanation. Please see Page 6, Lines 214, 218.

 

Point 8: Why did the authors decided to take a first threshold-based layer to make the distinction between lifts and other types? A general ML-based approach would not be more simple?

 

Response 8: Thank you for your valuable suggestion. Yes, the ML method is fully usable and logically simpler. However, since the data characteristics of static and elevator movement up and down are more obvious, they can be distinguished well even without ML, and this can save computing resources, so we used the method in the manuscript. At the same time, the data characteristics of several activities in Figure 2 may be more intuitive for viewing this feature. Please see Pages 6, Lines 219-224.

 

Point 9: In 2.4, the authors refer to the feature vector as the values of one single computation/feature. But the feature vector is the aggregate of all features.

 

Response 9: Thank you for your valuable suggestion. Yes, the feature vector is the aggregate of all features. We have revised the description in this section. Please see Page 6, Lines 230-234; Page 16, Line 545, etc.

 

Point 10: I do not understand how the authors obtained the 88 features if there were 28 features in 8 dimensions plus 3 more for the barometric data.

 

Response 10: Thank you for your valuable suggestion. It may be that the description of Table 1 was not clear enough. Take "standard deviation" as an example, calculate based on eight kinds of data, and generate eight characteristic values. The eight types of data include: 3-axis accelerometer data before and after filtering, and total accelerometer data before and after filtering. Not all algorithms use these eight types of data to calculate and generate feature values. Some may use only one type, and some use three. In the end, 28 algorithms used different types of data for the calculation, and we finally generated 88 features. We also have adjusted the format and contents of Table 1, hoping to express ourselves more clearly. Please see Page 6, Table 1 on Line 234.

 

Point 11: Table 1 is confuse. It should have one feature per line, maybe using two sided tables for space optimization.

 

Response 11: Thank you for your valuable suggestion. The description of Table 1 is not clear enough. We have adjusted the format and content of Table 1, hoping to express ourselves more clearly. The table distributes all 31 algorithms in nine types of data according to the data situation, and finally displays all 91 feature values. Please see Page 6, Table 1 on Line 234.

 

Point 12: The authors say that the limitation of the literature are related to the use of Wi-Fi and barometer. But the authors also use the Wi-Fi for first floor initiation and the barometer for feature extraction. How do the authors surpass the limitations of the literature then?

 

Response 12: Thank you for your valuable suggestion. In this paper, the use of air pressure data as a classification feature is beneficial to distinguish between up and down stairs and plane walking motion, but does not simply use air pressure data to determine floors because only relying on air pressure to judge floors has more limitations and lower universality. We have made corresponding adjustments and given detailed descriptions in the introduction. The Wi-Fi-based floor positioning accuracy is high, and can be used as the initial floor input, but its stability is inadequate. The method proposed in this paper is stable, and the two can be integrated with each other to achieve better floor positioning effects. For the description based on the Wi-Fi method, we have also made a comprehensive revision in the introduction. In addition, the focus of this article was to highlight the relative accuracy and stability of the HAR-based method. Please see Page 2, Lines 61-80 and Page 3, Lines 126-131, etc.

 

Point 13: Also, the authors say that they will address the real-time detection, but in Section 3 they say that the floor positioning is done recursively.

 

Response 13: Thank you for your valuable suggestion. The recursive process of floor positioning in this manuscript is embodied in the processing of each step in the stairwell, which can reflect the transition between floors. Most methods only detect the moment when the floor change is successful. Relatively speaking, the method in this paper has a better real-time performance. Please see Page 16, Figure 11 for the realization effect.

 

Point 14: Futhermore, the authors say that the limitations of the literature are related with the need for maintenance and implementation, which harms scalability. But this system also requires an extensive detail of all floor plans.

 

Response 14: Thank you for your valuable suggestion. The maintenance here refers to the maintenance of the fingerprint data or AP deployment environment in the wireless signal method. The fingerprint library collection workload is large. The method in this article is equivalent to obtaining the attributes of the stairs at only once time. The internal stairs structure of the building is more stable than the wireless signal deployment conditions. In most cases, the stair structure in a multi-storey building remains unchanged. At the same time, we have made a lot of revisions and improvements to the description in this part of the introduction. Please see Page 2, Lines 61-80, etc.

 

Point 15: Table too also has parameters that are not explained.

 

Response 15: Thank you for your valuable suggestion. We have made corresponding supplements and improvements. Please see Page 5, Line 194 and Page 9, Lines 289-295.

 

Point 16: How many data points were used for the linear regression? Isn't the model overfitted to a single elevator and device?

 

Response 16: Thank you for your valuable suggestion. We took the elevator more than 20 times in the five-story test site, and generated the above-mentioned elevator and height relationship diagram. For any floor position, the elevator running height value is discrete and discontinuous, and the running height difference between different floors is at least 4 meters. The model in the manuscript can reflect the relationship between the elevator running height and the running time, and can accurately identify the relative floor changes of the elevator running. Please see Page 8, Table 2 on Line 274, the column ‘floor_h’ in Table 2.

 

Point 17: What is the meaning of "temp" table?

 

Response 17: Thank you for your valuable suggestion. We have modified and improved the description of the table. This table is mainly used for the cumulative storage of certain movement status data of pedestrians, which can assist in the judgment of floor changes. Please see Page 9, Lines 289, 291-295.

 

Point 18: What is the 'Activity results' on Figure 8?

 

Response 18: Thank you for your valuable suggestion. We changed the name of this legend, which reflects the pedestrian activity category, and the red dot represents walking. Please see Page 14, Figure 9.

 

Point 19: How do the authors assure that the system is not in general overfitted? If more than 3 consecutive misclassification happen, the whole system will fail.

 

Response 19: Thank you for your valuable suggestion. The HAR-based floor change judgment method relies on the result of the HAR, and the 3-step threshold is also specified based on the test result of the HAR. In response to this situation, we increased the number of testers in the activity classification test to prove the rationality of this threshold. (Please see Page 13, Lines 414-434.) In addition, if there are more than three consecutive misclassifications, after the pedestrians' real floor switch is successful, it can be returned to the right floor location in time through the right walking state instead of system failure. At the same time, there is an explanation at the end of the article. (Please see Page 16, Lines 543-547.) At this time, we can improve the training model by adding training samples, or adjust the classification features to optimize it according to the data situation. If the maximum number of consecutive misjudgement steps is still four steps and there is no improvement, we can adjust the threshold in the floor change process to increase the fault tolerance rate and increase the decision delay to achieve floor change detection.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposed an algorithm for detection of floor evaluations based on the users’ walking statuses (defined as a type of the human activity recognition, HAR) predicted from a trained model and the observations captured by smartphone’s accelerometers. The algorithm starts with an accuracy ground initial evaluation and gradually detected the change of the evaluation after prediction of the users’ motions (still, plane walking, up/down stairs) at each location.

This is an interesting paper, but the presentation requires polishes. There are several major issues:

• The performance of the algorithm greatly relies on the three-axis accelerometer data(TAAD). However, throughout the entire paper, I cannot see any theoretical discussions or experiments related to the basic accelerometer’s error model which includes the bias, scale factor, cross-axis coupling factors, and etc. The authors should at least include a small section to discuss how the accelerometer’s errors affect the final results (as claimed nearly 100%).
• The authors showed more than 30 ways to analyze the feature vectors, but they are not used (not compared) to deliver different results. The paper has not shown any evidence to explain which tools/statistics should be used to recognize the human activities. How do the different choices affect the final accuracy?

There are many English presentation issues, please check clearly the English before re-submit the paper again? There are some issues: 

1) in the abstract: Line 18, “whether pedestrians had changed their floor positions”- I think the floor position can be never changed. The pedestrians can only change their motions on a floor, and the floor’s evaluation cannot be changed.

2) The last sentence of the abstract is very unclear.

3) Line 98 “The method is not sensitive to changes of indoor structure…” So the method is not sensitive to the ground’s evaluation? If so, how could the method achieve 100% accuracy?

4) Line 103, “change of floor”, again the floor does not change, only it’s elevation changes.

5) Figure 1, I do not see why it’s necessary to rotate 90 degree for some important steps of the flow, please correct it.

6) The code after 157 is not a standardized pseudo-code representation.

7) In Table 1, why “Kurtosis” is not “Kurtosis coefficient” if ““Skewness coefficient” is used later on?, why the author had to listed 31 features vectors as there are total 88 (Line 187)

8) Why Figure 6 is mentioned in Line 195? but Figure 6 appears after Line 361. Figure 6 is very difficult to understand, what the are x – y axes of the figure stand for, what does the red color stand for?

9) From Line 240 to 301, there are Algorithms 1 to 5 (Algorithm 3.1,3.2,3.3), are there so many algorithms in the paper?? You might want to use “Step 1” of “Algorithm 1”

10) Figures 9 and 10 are important figures showing the results but the resolution is very low. Please increase the resolution.

Author Response

Response to Reviewer 2 Comments

 

This paper proposed an algorithm for detection of floor evaluations based on the users’ walking statuses (defined as a type of the human activity recognition, HAR) predicted from a trained model and the observations captured by smartphone’s accelerometers. The algorithm starts with an accuracy ground initial evaluation and gradually detected the change of the evaluation after prediction of the users’ motions (still, plane walking, up/down stairs) at each location.

This is an interesting paper, but the presentation requires polishes. There are several major issues:

 

Point 1: The performance of the algorithm greatly relies on the three-axis accelerometer data (TAAD). However, throughout the entire paper, I cannot see any theoretical discussions or experiments related to the basic accelerometer’s error model which includes the bias, scale factor, cross-axis coupling factors, and etc. The authors should at least include a small section to discuss how the accelerometer’s errors affect the final results (as claimed nearly 100%).

 

Response 1: Thank you for your valuable suggestion. It is a recognized fact that there are errors in smartphone accelerometers. Similar to most documents, we used mean filtering to eliminate the effect of "noise"; at the same time, in the HAR classification stage, most of the proposed feature vectors selected filtered data, as shown in Table 2 of the revised manuscript, which can also effectively avoid the error of acceleration data. In addition, the experiment with the lightest color in the white dashed box represents the lowest classification accuracy in Figure 3, and was performed with most part of unfiltered accelerometer data. The accuracy of this result is too low to be used at all. The filtered eigenvalues can effectively shield the influence of "noise" data, and lead to a higher classification accuracy. The relevant description of this point has been added in Section 2.5. Please see Page 4, Lines 173-175and Page 7, Lines 247-252.

 

Point 2: The authors showed more than 30 ways to analyze the feature vectors, but they are not used (not compared) to deliver different results. The paper has not shown any evidence to explain which tools/statistics should be used to recognize the human activities. How do the different choices affect the final accuracy?

 

Response 2: Thank you for your valuable suggestion. Our description in the original manuscript was not clear enough. It has been improved in Figure 3 and supplemented with related algorithms. We have indicated that the columns represent various feature vectors. Please see Page 7, Lines 236-244 and Page 8, Figure 3.

 

Point 3: There are many English presentation issues, please check clearly the English before re-submit the paper again? There are some issues:

Point 3.1: in the abstract: Line 18, “whether pedestrians had changed their floor positions”- I think the floor position can be never changed. The pedestrians can only change their motions on a floor, and the floor’s evaluation cannot be changed.

 

Response 3.1: Thank you for your valuable suggestion. We revised the description and have revised the manuscript through a professional editing service. Please see Page 1, Line 24.

 

Point 3.2: The last sentence of the abstract is very unclear.

 

Response 3.2: Thank you for your valuable suggestion. We have revised the manuscript through a professional editing service. Please see Page 1, Line 29-30.

 

Point 3.3: Line 98 “The method is not sensitive to changes of indoor structure…” So the method is not sensitive to the ground’s evaluation? If so, how could the method achieve 100% accuracy?

 

Response 3.3: Thank you for your valuable suggestion. This method can be adapted to a variety of multi-floor indoor structures. That is to say, if you change to another multi-floor environment, even if the floor number is different and the floor height is different, the HAR-based method can still be used based on the staircase reference data in the new multi-floor environment, to achieve an ideal floor positioning effect. The wireless signal floor recognition method is likely to be affected by problems such as the AP layout and changes in the internal spatial structure, leading to a decrease in accuracy. Please see Page 3, Line 126-131.

 

Point 3.4: Line 103, “change of floor”, again the floor does not change, only it’s elevation changes.

 

Response 3.4: Thank you for your valuable suggestion. We have revised the description. Please see Page 3, Line 138.

 

Point 3.5: Figure 1, I do not see why it’s necessary to rotate 90 degree for some important steps of the flow, please correct it.

 

Response 3.5: Thank you for your valuable suggestion. We have corrected it in accordance with your comments. Please see Page 4, Line 168.

 

Point 3.6: The code after 157 is not a standardized pseudo-code representation.

 

Response 3.6: Thank you for your valuable suggestion. We have changed these codes to a standardized pseudo-code. Please see Page 5, Line 194.

 

Point 3.7: In Table 1, why “Kurtosis” is not “Kurtosis coefficient” if ““Skewness coefficient” is used later on?, why the author had to listed 31 features vectors as there are total 88 (Line 187).

 

Response 3.7: Thank you for your valuable suggestion. We have changed “Kurtosis” to “Kurtosis coefficient”, and we adjusted the format and contents of Table 1. Please see Page 7, Line 234, row 25 in Table 1.

 

Point 3.8: Why Figure 6 is mentioned in Line 195? but Figure 6 appears after Line 361. Figure 6 is very difficult to understand, what the are x – y axes of the figure stand for, what does the red color stand for?

 

Response 3.8: Thank you for your valuable suggestion. We put Figure 6 to the front, and it is now Figure 3. In the figure, the x row represents different classification algorithms, and the y column represents different feature value combinations. Red is used to indicate the classification accuracy of the test data. The higher the accuracy, the darker the color. The description and explanation are given in the figure and before the figure. Please see Page 7, Lines 236-244 and Figure 3 in Page 8.

 

Point 3.9: From Line 240 to 301, there are Algorithms 1 to 5 (Algorithm 3.1,3.2,3.3), are there so many algorithms in the paper?? You might want to use “Step 1” of “Algorithm 1”.

 

Response 3.9: Thank you for your valuable suggestion. We have changed the content. The flowchart describes the overall processing flow, and the algorithm has been changed into steps. Please see Pages 9-11, Lines 297-373.

 

Point 3.10: Figures 9 and 10 are important figures showing the results but the resolution is very low. Please increase the resolution.

 

Response 3.10: Thank you for your valuable suggestion. We have modified the picture and increased the resolution. At the request of another reviewer, we have added an experiment to compare our method with other HAR-based methods, which is reflected in the addition of a curve in Figure 10. Please see Figure 10 on Page 15 and Figure 11 on Page 16.

 

Author Response File: Author Response.docx

Reviewer 3 Report

The paper describes a method able to define the floor positioning. The method is based on human activity recognition (HAR). The article is clear and well structured. The main drawback is in the comparison results. The authors compare the proposed method with too simplistic methods: barometric and WiFi. The comparison with methods that merge different information would be more fair.

Moreover, some recent works on indoor localization are missing. The introduction needs to clearly describe some aspects in order to show that indoor localization is an hot topic. For instance, the future direction of indoor localization systems ("What is next for Indoor Localisation? Taxonomy, protocols, and patterns for advanced Location Based Services"), how it is possible to fairly evaluate indoor localization systems ("The IPIN 2019 Indoor Localisation Competition—Description and Results", "Evaluating AAL solutions through competitive benchmarking: the localization competition"), standard, the issue of privacy&security, system integration, crowds localization.

These are just some references suggested, but the important is that these aspects are covered in the introduction. 

Author Response

Response to Reviewer 3 Comments

 

Point 1: The paper describes a method able to define the floor positioning. The method is based on human activity recognition (HAR). The article is clear and well structured. The main drawback is in the comparison results. The authors compare the proposed method with too simplistic methods: barometric and WiFi. The comparison with methods that merge different information would be more fair.

 

Response 1: Thank you for your valuable suggestion. On the basis of the original results, we added a HAR-based method to compare with the method in this article. The result is shown in Figure 10. At the same time, the description of the method was also added to the previous part of the figure. Please see Page 15, Lines 485, 491 and Figure 10.

 

Point 2: Moreover, some recent works on indoor localization are missing. The introduction needs to clearly describe some aspects in order to show that indoor localization is an hot topic. For instance, the future direction of indoor localization systems ("What is next for Indoor Localisation? Taxonomy, protocols, and patterns for advanced Location Based Services"), how it is possible to fairly evaluate indoor localization systems ("The IPIN 2019 Indoor Localisation Competition—Description and Results", "Evaluating AAL solutions through competitive benchmarking: the localization competition"), standard, the issue of privacy&security, system integration, crowds localization.

These are just some references suggested, but the important is that these aspects are covered in the introduction.

 

Response 2: Thank you for your valuable suggestion. We researched some of the latest works on indoor positioning, and added some explanations in the introduction, and made substantial revisions to further highlight the popular nature of indoor positioning and the theme of this article. Please see Page 1, Lines 35; Page 1, Lines 50-53, 61-63 and Page 15, Lines 498-499, etc. We also have added some corresponding references in the reference section.

Author Response File: Author Response.docx

Reviewer 4 Report

The research is conducted correctly with high standards of the process. The paper is written in a proper way in terms of expectations from scientific paper. Some language flaws can be found that requires correction.

The topic is not of a high importance, as many Indoor Positioning Systems detect the floor with satisfactory accuracy. This is true that some conditions may make the detection problematic and the paper showed that the system may benefit from HAR based inference.

Suggestions:

• Not consistent use of grouping digits in numbers: in most cases grouping 3 digits, but in line 369: "100, 000", line 371 "4200" - no grouping, etc...
• Chapter 2.5 seems to discussed classification algorithms, however, there is another chapter on classification - 4.3. It looks like inconsistency, there is even no reference between them.
• Line 208 - what is "floor ladder library"?
• Line 213 - multistory? rather multi-storey
• Figure 3. The diagram resembles UML Sequence Diagram, but there are some discrepancies from the standard. It would be good to follow the standard.
• Line 345: "More than ten classification algorithms were chosen, such as tree, Bayes, SVM, nearest neighbour, rule, KNN and etc" - these are not correct names of classificators: "tree" - did you mean "decision tree"?, "rule" - no idea?, "Bayes" - "Naive Bayes"?, what's the difference between "nearest neighbour" and "KNN"?
• Figure 6 - it is not explained in the text nor in the image description - what 12 algorithms were compared.
• Line 345-361 - 88 feature vectors suggest number of vectors rather than number of dimensions of the training vectors; PCA reduces number of dimensions, here from 88 to 18
• Figure 9 and 10 - blurred, illegible.

Author Response

Response to Reviewer 4 Comments

 

The research is conducted correctly with high standards of the process. The paper is written in a proper way in terms of expectations from scientific paper. Some language flaws can be found that requires correction.

The topic is not of a high importance, as many Indoor Positioning Systems detect the floor with satisfactory accuracy. This is true that some conditions may make the detection problematic and the paper showed that the system may benefit from HAR based inference.

Suggestions:

 

Point 1: Not consistent use of grouping digits in numbers: in most cases grouping 3 digits, but in line 369: "100, 000", line 371 "4200" - no grouping, etc...

 

Response 1: Thank you for your valuable suggestion. We have changed the way numbers are expressed. Please see Page 7, Line 254; Page 13, Line 378; and Page 13, Line 424.

 

Point 2: Chapter 2.5 seems to discussed classification algorithms, however, there is another chapter on classification - 4.3. It looks like inconsistency, there is even no reference between them.

 

Response 2: Thank you for your valuable suggestion. We have adjusted the chapter structure, the two subsections have been partially merged, and Figure 6 has been advanced to the position of Figure 3. Please see Page 7, Lines 236-260.

 

Point 3: Line 208 - what is "floor ladder library"?

 

Response 3: Thank you for your valuable suggestion. It means “floor stair library”, and it is used to save the corresponding number of stairs in each floor. Please see Page 8, Lines 271- 274.

 

Point 4: Line 213 - multistory? rather multi-storey

 

Response 4: Thank you for your valuable suggestion. We have uniformly changed “multistory” to “multi-storey” in the full manuscript. Please see Page 8, Line 271, 276, etc.

 

Point 5: Figure 3. The diagram resembles UML Sequence Diagram, but there are some discrepancies from the standard. It would be good to follow the standard.

 

Response 5: Thank you for your valuable suggestion. We have look over the standard format of UML sequence diagrams and found that it is slightly different from the solution we described. Most UML diagrams mainly describe the calculation logic of the algorithm unit, and our purpose was to describe the logical processing relationship between the different steps in the entire program. We considered various forms of diagrams and found that this expression is more vivid and easy to understand, so we did not make corresponding adjustments. At the same time, at the request of another reviewer, we uniformly changed the “algorithm” in the article to “step”. Please see Page 10, Line 306.

 

Point 6: Line 345: "More than ten classification algorithms were chosen, such as tree, Bayes, SVM, nearest neighbour, rule, KNN and etc" - these are not correct names of classificators: "tree" - did you mean "decision tree"?, "rule" - no idea?, "Bayes" - "Naive Bayes"?, what's the difference between "nearest neighbour" and "KNN"?

 

Response 6: Thank you for your valuable suggestion. We apologize that the description of each algorithm was not accurate enough due to the two software programs used in the classification test process. The "tree" means "decision tree", "rule" means the "Decision Table" algorithm in the Weka software, "Bayes" means "Naive Bayes" and "Bayes Net" algorithm. "Nearest neighbour" means "KNN". We have added the names of the algorithms to Figure 3, and adjusted and improved the description of this part in the manuscript. Please see Page 7, Lines 236-238, and Figure 3 on Page 8.

 

Point 7: Figure 6 - it is not explained in the text nor in the image description - what 12 algorithms were compared.

 

Response 7: Thank you for your valuable suggestion. We have added the corresponding algorithm name to the manuscript and the image. Page 7, Lines 236-238, and Figure 3 on Page 8.

 

Point 8: Line 345-361 - 88 feature vectors suggest number of vectors rather than number of dimensions of the training vectors; PCA reduces number of dimensions, here from 88 to 18

 

Response 8: Thank you for your valuable suggestion. We have made adjustments to the description and revised the format of Table 1. Please see Page 6, Lines 229-234 and Table 1. In addition, the PCA method is a general method that uses linear transformation to transform multiple variables into a few variables with the same classification effect. The PCA method can save computing resources and is a commonly used method. In the manuscript, we have integrated the original 2.5 and 4.3 sections. For the PCA method, we do not involve any innovation, and the classification accuracy with the two variables is the same, so we did not introduce it.

 

Point 9: Figure 9 and 10 - blurred, illegible.

 

Response 9: Thank you for your valuable suggestion. We have modified the picture and increased the resolution. At the request of another reviewer, we have added an experiment to compare our method with other HAR-based methods, which is reflected in the addition of a curve in Figure 10. Please see Page 15, Figure 10 and Page 16, Figure 11.

Author Response File: Author Response.docx