Classifying Participant Standing and Sitting Postures Using Channel State Information

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The work of this manuscript has a certain contribution to passive HAR based on CSI signals. My concerns include:

 

1. The authors collected a (small) dataset and proposed an improved model. The title seems to emphasize the method rather than the dataset. The main problem is that only one-way validation was used. The authors experimented with the proposed method on their data, as well as two external methods for comparison. This is good. But what needs to be supplemented is to apply their method to other, e.g., open-source datasets, and compare the results with existing benchmarks/SOTAs. This is correct two-way validation for a rigorous scientific paper.

Many CSI-based HAR datasets exist, such as StanWiFi. When the authors apply the proposed method to them, they may encounter the following two issues, both of which, nevertheless, are effortless to solve:

a. Irrelevant activities in the dataset. Almost all CSI HAR datasets involve the two primary activities of standing and sitting, of which the data portion can be taken out with the others eliminated.

b. More than one antenna. No matter how many antennas there are, they are essentially measuring a specific activity simultaneously. Assuming that a dataset uses three pairs of transmitting and receiving antennas, you only need to treat each receiving antenna as an independent signal to test your method.

 

At least certain external data needs to be used to prove the applicability, generalizability, and universality of your method. Otherwise, it is solely a work to deliberately collect highly targeted datasets just to reflect the design of a certain algorithm, ignoring the reuse value of the method—In this case, the manuscript must be changed to a claim that a small-scale (specific purpose) dataset was collected (and made public) and a method was implemented on it. Nothing more. In this case, it seems to be just a conference paper, but far from a full-fledged journal article.

 

2. The authors did not explain the reasons for selecting some parameters. For example, why was the collection duration of 50 seconds (25+25) chosen?

 

3. The narrative in Section 2.1 is weak and far from the headings. In terms of sensor-based HAR, the authors only briefed external sensing, but entirely ignored internal sensing (mainly wearables). For example, mobile phones that people carry daily and wearable HAR-related sensors integrated into devices that specific groups of people must wear (such as knee bandages)... these technologies have the characteristics of non-invasiveness, privacy, wide range of mobility, low cost, and high accuracy. What are the advantages of Wi-Fi compared to wearables? The recent SOTA of INTERNAL sensing for HAR include On a real real-time wearable human activity recognition system; as well as High-level features for human activity recognition and modeling.

 

Moreover, the most recent literature that is the most suitable for the title of Section 2.1 is Sensor-based human activity and behavior research. It concluded: "High-quality sensory observations applicable to recognizing users' activities and behaviors, including electrical, magnetic, mechanical (kinetic), optical, acoustic, thermal, and chemical biosignals, are inseparable from sensors' sophisticated design and appropriate application... Traditional sensors suitable for HAR, including external sensors for smart homes, optical sensors such as cameras for capturing video signals, and bioelectrical, biomagnetic, and biomechanical sensors for wearable applications, have been studied and verified adequately... Meanwhile, innovative sensor research for HAR or HBR is also very active in the academic community, including new sensors appropriate for HAR/HBR, new designs and applications of the above-mentioned traditional sensors, and the usage of non-traditional HAR/HBR-related sensor types, among others." It can also be seen from these words that the current 2.1 is very incomplete.

 

4. Figure 3 is unacceptable. Try simply inverting the color and adjusting it.

 

5. The coloring of Fig.5 is wrong. It's just that because the number of instances of each class is relatively close in your case, there appears to be problem-free. The coloring of the confusion matrix should be based on percentages rather than instance quantities (why? Imagine a situation where you make a binary classification of standing and sitting, and the results are all correct. There are 200 samples in standing and 100 in sitting. According to your current coloring and color map, what will be the plot?) In addition, it is recommended to display two lines in each cell; first the quantity, and second the percentage — refer to this most recent work: Efficient Wi-Fi-Based Human Activity Recognition Using Adaptive Antenna Elimination. By the way, for CSI-based HAR, this article is also the latest SOTA ignored by the author. It studies the case (including standing and sitting) of multi-antenna HAR and obtains very eye-catching but low-cost ML results.

Author Response

We would like to thank the reviewer for their useful comments, which have helped us improve our manuscript. Please find our responses below.

1. The authors collected a (small) dataset and proposed an improved model. The title seems to emphasize the method rather than the dataset. The main problem is that only one-way validation was used. The authors experimented with the proposed method on their data, as well as two external methods for comparison. This is good. But what needs to be supplemented is to apply their method to other, e.g., open-source datasets, and compare the results with existing benchmarks/SOTAs. This is correct two-way validation for a rigorous scientific paper.

 

Many CSI-based HAR datasets exist, such as StanWiFi. When the authors apply the proposed method to them, they may encounter the following two issues, both of which, nevertheless, are effortless to solve:

 

1. Irrelevant activities in the dataset. Almost all CSI HAR datasets involve the two primary activities of standing and sitting, of which the data portion can be taken out with the others eliminated.

 

1. More than one antenna. No matter how many antennas there are, they are essentially measuring a specific activity simultaneously. Assuming that a dataset uses three pairs of transmitting and receiving antennas, you only need to treat each receiving antenna as an independent signal to test your method.

 

At least certain external data needs to be used to prove the applicability, generalizability, and universality of your method. Otherwise, it is solely a work to deliberately collect highly targeted datasets just to reflect the design of a certain algorithm, ignoring the reuse value of the method—In this case, the manuscript must be changed to a claim that a small-scale (specific purpose) dataset was collected (and made public) and a method was implemented on it. Nothing more. In this case, it seems to be just a conference paper, but far from a full-fledged journal article.

 

Author response: We have used an additional publicly available dataset which performs two relevant activities to our paper, sitting and standing. We performed our tested models, in particular the LSTM-1DCNN, to test its capabilities on unseen data. We also evaluated our results against other models which had utilised the same dataset for comparison. Unfortunately, we couldn’t access the StanWiFi dataset after reviewing their paper and contacting the authors, but the publicly available one we used is appropriate.

 

2. The authors did not explain the reasons for selecting some parameters. For example, why was the collection duration of 50 seconds (25+25) chosen?

 

Author response: We have added more detail to this section as to the logic behind selecting 50 seconds of sampling time. In addition to this, we have also explained our reasoning for testing the three models with adequate detail.

 

 

3. The narrative in Section 2.1 is weak and far from the headings. In terms of sensor-based HAR, the authors only briefed external sensing, but entirely ignored internal sensing (mainly wearables). For example, mobile phones that people carry daily and wearable HAR-related sensors integrated into devices that specific groups of people must wear (such as knee bandages)... these technologies have the characteristics of non-invasiveness, privacy, wide range of mobility, low cost, and high accuracy. What are the advantages of Wi-Fi compared to wearables? The recent SOTA of INTERNAL sensing for HAR include On a real real-time wearable human activity recognition system; as well as High-level features for human activity recognition and modeling.

 

 Moreover, the most recent literature that is the most suitable for the title of Section 2.1 is Sensor-based human activity and behavior research. It concluded: "High-quality sensory observations applicable to recognizing users' activities and behaviors, including electrical, magnetic, mechanical (kinetic), optical, acoustic, thermal, and chemical biosignals, are inseparable from sensors' sophisticated design and appropriate application... Traditional sensors suitable for HAR, including external sensors for smart homes, optical sensors such as cameras for capturing video signals, and bioelectrical, biomagnetic, and biomechanical sensors for wearable applications, have been studied and verified adequately... Meanwhile, innovative sensor research for HAR or HBR is also very active in the academic community, including new sensors appropriate for HAR/HBR, new designs and applications of the above-mentioned traditional sensors, and the usage of non-traditional HAR/HBR-related sensor types, among others." It can also be seen from these words that the current 2.1 is very incomplete.

 

Author response: We reviewed this section and expanded on it in greater detail, analysing recent literature as well as from the previous year and elaborating in great detail on why WiFi sensing is better than vision or wearable sensor-based.

 

4. Figure 3 is unacceptable. Try simply inverting the color and adjusting it.

 

Author response: The figure has now been inverted to make it a white background.

 

5. The coloring of Fig.5 is wrong. It's just that because the number of instances of each class is relatively close in your case, there appears to be problem-free. The coloring of the confusion matrix should be based on percentages rather than instance quantities (why? Imagine a situation where you make a binary classification of standing and sitting, and the results are all correct. There are 200 samples in standing and 100 in sitting. According to your current coloring and color map, what will be the plot?) In addition, it is recommended to display two lines in each cell; first the quantity, and second the percentage — refer to this most recent work: Efficient Wi-Fi-Based Human Activity Recognition Using Adaptive Antenna Elimination. By the way, for CSI-based HAR, this article is also the latest SOTA ignored by the author. It studies the case (including standing and sitting) of multi-antenna HAR and obtains very eye-catching but low-cost ML results.

 

Author response: All confusion matrices have now been redone in response to your point above. The colour is now based on the percentages rather than instance quantities. Thank you for raising this point.

Reviewer 2 Report

Comments and Suggestions for Authors

In this manuscript, the use of CSI data obtained from an ESP32 microcontroller to identify participants from sitting and standing postures has been investigated in a many-to-one classification method. Various pre-processing algorithms and techniques have been incorporated and tested on different classification algorithms. Performance metrics were captured such as the confusion matrix, accuracy and elapsed time. The hybrid LSTM-1DCNN model has an average accuracy of 84.29% and 74.13% for sitting and standing postures. Results show that this model is with higher efficiency. 

This reviewer thinks the research is significant. However, the writing must be improved. Such as, Fig.2 is not clear for expression of author’s ideas. Fig. 3 is too dark to read nothing.

Author Response

We would like to thank the reviewer for their useful comments, which have helped us improve our manuscript. Please find our responses below.

In this manuscript, the use of CSI data obtained from an ESP32 microcontroller to identify participants from sitting and standing postures has been investigated in a many-to-one classification method. Various pre-processing algorithms and techniques have been incorporated and tested on different classification algorithms. Performance metrics were captured such as the confusion matrix, accuracy and elapsed time. The hybrid LSTM-1DCNN model has an average accuracy of 84.29% and 74.13% for sitting and standing postures. Results show that this model is with higher efficiency.

 

This reviewer thinks the research is significant. However, the writing must be improved. Such as, Fig.2 is not clear for expression of author’s ideas. Fig. 3 is too dark to read nothing.

 

Author response: Figure 3 colour has now been inverted to a white background making the figure much clearer to see. However, we believe that Figure 2 portrays exactly what the methodology is, and the steps involved. We have also improved the text throughout the manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript titled "Classifying Participant Standing and Sitting Postures Using Channel State Information" presents a comprehensive study leveraging Channel State Information (CSI) for human activity recognition, particularly distinguishing between sitting and standing postures. The authors have embarked on a meticulous journey to explore this niche, demonstrating the efficacy of CSI data captured via an ESP32 microcontroller in a controlled environment. The methodology is robust, with a commendable effort in integrating various pre-processing algorithms and testing them on different classification paradigms.

However, despite the strengths of the paper, there are several areas where minor revisions could significantly enhance the manuscript's quality and contribution to the field.

·         Sample Size and Diversity: While the authors have managed to recruit 15 participants for the study, the sample size is relatively constrained given the variability in human postures and the potential diversity in environmental contexts. Expanding the participant base would not only enhance the generalizability of the findings but also provide a richer dataset for algorithm training. Furthermore, the inclusion of participants with varying physical characteristics and from different demographic backgrounds could introduce additional dimensions to the study.

·         Comparative Analysis: The paper would benefit substantially from a more extensive comparative analysis. While the hybrid LSTM-1DCNN model shows promise, situating these results within the broader landscape of similar research would provide readers with a clearer understanding of the study's innovations or superiority. It is advisable for the authors to delve into parallel studies, comparing methodologies, and performance metrics, which would position their contributions within the ongoing scholarly dialogue.

·         Environmental Variables: The study's reliance on a controlled isolated environment, while beneficial for standardization, raises questions about the model's performance in more chaotic, real-world scenarios. Future iterations of this research could address these concerns by perhaps introducing controlled environmental variables or conducting pilot tests in varied settings.

·         Model Interpretability and Parameters: The manuscript could articulate more clearly the interpretability of the models used, especially the hybrid LSTM-1DCNN model. While the technical aspects are profoundly addressed, a layman's summary or a more in-depth discussion on why certain parameter selections were made would make the paper more accessible to a broader audience, including those outside the immediate field of study.

·         Ethical Considerations: Given the nature of human activity recognition, a brief discourse on the ethical implications of this technology, privacy considerations, and data security would provide a more holistic view of the field's trajectory. This discussion is pertinent given the increasing integration of such technologies in sensitive areas such as patient monitoring and personal fitness.

In conclusion, the paper is a valuable contribution to the field of human activity recognition using CSI. The minor concerns outlined above notwithstanding, the research lays significant groundwork for future exploration. With slight refinements, particularly in expanding the comparative analysis and addressing the model's real-world applicability, the manuscript stands as a strong candidate for publication. The authors are to be commended for their rigorous approach and clear exposition of a complex subject matter.

Author Response

We would like to thank the reviewer for their useful comments, which have helped us improve our manuscript. Please find our responses below.

However, despite the strengths of the paper, there are several areas where minor revisions could significantly enhance the manuscript's quality and contribution to the field.

 

Sample Size and Diversity: While the authors have managed to recruit 15 participants for the study, the sample size is relatively constrained given the variability in human postures and the potential diversity in environmental contexts. Expanding the participant base would not only enhance the generalizability of the findings but also provide a richer dataset for algorithm training. Furthermore, the inclusion of participants with varying physical characteristics and from different demographic backgrounds could introduce additional dimensions to the study.

 

Author response: Unfortunately, due to a short turnaround we simply don’t have enough time to gather more participants due to the many steps involved with setting up and recruiting etc. However, we performed a two-way validation to show that the models can generalise well to different datasets.

 

 

Comparative Analysis: The paper would benefit substantially from a more extensive comparative analysis. While the hybrid LSTM-1DCNN model shows promise, situating these results within the broader landscape of similar research would provide readers with a clearer understanding of the study's innovations or superiority. It is advisable for the authors to delve into parallel studies, comparing methodologies, and performance metrics, which would position their contributions within the ongoing scholarly dialogue.

 

Author response: We have now gone and found a publicly available dataset which performs two relevant activities to our paper, sitting and standing. We used our tested model, the LSTM-1DCNN, to test its capabilities on unseen data. We also evaluated our results against other models which had utilised the same dataset for comparison.

 

Environmental Variables: The study's reliance on a controlled isolated environment, while beneficial for standardization, raises questions about the model's performance in more chaotic, real-world scenarios. Future iterations of this research could address these concerns by perhaps introducing controlled environmental variables or conducting pilot tests in varied settings.

 

Author response: We decided to add a brief paragraph explaining this limitation and linked it to a future exploration of this.

 

 

Model Interpretability and Parameters: The manuscript could articulate more clearly the interpretability of the models used, especially the hybrid LSTM-1DCNN model. While the technical aspects are profoundly addressed, a layman's summary or a more in-depth discussion on why certain parameter selections were made would make the paper more accessible to a broader audience, including those outside the immediate field of study.

 

Author response: We decided to add a lot more detail on the reasoning behind the selection of all 3 models, and in particular LSTM-1DCNN to give the reader a clear understanding of the numerous benefits of each model should they decide to use them.

 

 

Ethical Considerations: Given the nature of human activity recognition, a brief discourse on the ethical implications of this technology, privacy considerations, and data security would provide a more holistic view of the field's trajectory. This discussion is pertinent given the increasing integration of such technologies in sensitive areas such as patient monitoring and personal fitness.

 

Author response: We added some detail concerning the ethical considerations regarding WiFi sensing.

 

 

In conclusion, the paper is a valuable contribution to the field of human activity recognition using CSI. The minor concerns outlined above notwithstanding, the research lays significant groundwork for future exploration. With slight refinements, particularly in expanding the comparative analysis and addressing the model's real-world applicability, the manuscript stands as a strong candidate for publication. The authors are to be commended for their rigorous approach and clear exposition of a complex subject matter.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors addressed most of my concerns. Especially, they updated the confusion matrices to a correct and high-level level. One thing that needs to be supplemented: the authors generated detailed and accurate confusion matrices on the inspiration of the recommended literature (https://doi.org/10.1109/ACCESS.2023.3320069) in review round 1, which should be referred to. It is also one of the few recent SOTAs utilizing CSI to identify sitting and standing effectively.

Because I like this work well, and in order to save the time and academic resources of the authors and editorial office, I don't insist on the need for another round of reviewing. I argue that the manuscript can be published after the above issue is resolved.

Author Response

Thank you for your positive comments and reviewing so quickly, it is greatly appreciated. We have resolved your concern by adding a reference to the works. We have specifically introduced the following statement: "The use of confusion matrices in this work was inspired by recent work published by Jannat et al.~\cite{jannat23}." in the Evaluation Section.