Efficient Transmission-Based Human Behavior Recognition Algorithm

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript electronics-3566478 proposes the KCS algorithm, a compressed sensing method based on K-SVD, to reduce the transmission burden of Channel State Information (CSI) data while maintaining high recognition accuracy. The key innovation lies in replacing traditional sparse matrices with an overcomplete sparse matrix, improving sparsity and compression efficiency. The authors claim a 90% reduction in data volume while achieving 90% accuracy in behavior recognition. While the compression method is well-motivated for real-time sensing systems, the paper lacks critical comparisons with existing deep learning-based approaches, such as Human Behavior Recognition Based on Multiscale Convolutional Neural Network, which employs multi-scale spatiotemporal feature extraction. The evaluation is limited to accuracy and compression rate, omitting computational efficiency, robustness in noisy environments, and generalization across datasets. The KCS algorithm shows promise but requires deeper ablation studies, parameter sensitivity analysis, and broader benchmarking to establish its superiority over existing methods. It was a pleasure reviewing this work and I can recommend it for publication in Electronics after a major revision. I respectfully refer the authors to my comments below.

 

Reviewer Comments

1. The English needs to be revised throughout. The authors should pay attention to the spelling and grammar throughout this work. I would only respectfully recommend that the authors perform this revision or seek the help of someone who can aid the authors.
2. (Section 1 Introduction) The reviewer hopes the introduction section in this paper can introduce more studies in recent years. The reviewer suggests authors don't list a lot of related tasks directly. It is better to select some representative and related literature or models to introduce with certain logic. For example, the latter model is an improvement on one aspect of the former model.
3. The proposed KCS algorithm focuses on data compression, while Human Behavior Recognition Based on Multiscale Convolutional Neural Network emphasizes spatiotemporal feature learning. These are complementary rather than competing approaches. The paper should: Compare end-to-end recognition accuracy when KCS is paired with a simple classifier (e.g., SVM) vs. the multiscale CNN. Discuss whether compression sacrifices discriminative features that deep learning models rely on. Justify why KCS is better than random sampling or PCA-based compression if it is purely a preprocessing step.
4. While the paper highlights 90% data reduction, it neglects: Latency vs. accuracy trade-off: How much does compression delay real-time prediction? Robustness to noise: Wireless CSI is often noisy—does KCS degrade under low SNR? Generalization: Test on multiple public datasets (e.g., UTD-MHAD, Berkeley MHAD) rather than a proprietary setup.
5. (Section I, Introduction) The reviewer suggest to revise the original statement as …Prior to the input of the reconstructed CSI data into the CNN classifier*, it is necessary to undertake preprocessing., ... some related work such as, orientation cues-aware facial relationship representation for head pose estimation via transformer; transifc: invariant cues-aware feature concentration learning for efficient fine-grained bird image classification; mmatrans: muscle movement aware representation learning for facial expression recognition via transformers.
6. The KCS algorithm relies on sparsity level, dictionary size (K), and compression ratio, but their impact is not analyzed. For example: Does a larger K in K-SVD improve recognition but increase computation? Is there an optimal compression ratio beyond which accuracy drops sharply? How does overcomplete matrix design compare to learned dictionaries (e.g., via autoencoders)?
7. Recent works (e.g., neural compression, autoencoders) achieve high compression with minimal feature loss. The paper should: Compare KCS with an LSTM-autoencoder for CSI sequence compression. Benchmark against quantization/pruning techniques used in edge AI.
8. The original sentence is suggested to revise as … Here, the CNN is used to avoid complicated feature extraction steps and has strong recognition and classification ability* , ... some related work such as, ldcnet: limb direction cues-aware network for flexible human pose estimation in industrial behavioral biometrics systems; ehpe: skeleton cues-based gaussian coordinate encoding for efficient human pose estimation; arhpe: asymmetric relation-aware representation learning for head pose estimation in industrial human-computer interaction.
9. The paper mentions KCS ensures data security, but no details are provided. For example: Does the sparse representation prevent eavesdropping or adversarial attacks? How does it compare to encrypted transmission (e.g., homomorphic encryption)?
10. The multiscale CNN paper discusses model compactness and transferability, whereas KCS ignores: Hardware constraints: Can KCS run on low-power IoT devices? Scalability: How does compression time scale with longer CSI sequences?
11. The multiscale CNN paper validates perceptual quality via feature visualizations. KCS should: Show reconstructed CSI waveforms to assess fidelity loss. Include failure cases (e.g., misclassified behaviors due to over-compression).
12. The title suggests behavior prediction, but the experiments focus only on static recognition. The paper should: Test temporal prediction (e.g., forecasting next action). Compare with RNN/LSTM-based methods for sequential behavior modeling.
13. Discuss the pros and cons of the proposed human behavior recognition models.

 

My overall impression of this manuscript is that it is in general well-organized. The work seems interesting and the technical contributions are solid. I would like to check the revised manuscript again.

Comments on the Quality of English Language

The English needs to be revised throughout. The authors should pay attention to the spelling and grammar throughout this work. I would only respectfully recommend that the authors perform this revision or seek the help of someone who can aid the authors.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript offers a significant advancement in the domain of human activity recognition through the utilization of Wi-Fi Channel State Information (CSI) data, emphasizing data compression and effective transmission via a newly proposed K-Singular Value Decomposition Compressed Sensing (KCS) framework. The research capitalizes on the principles of compressed sensing and sparse dictionary learning to minimize the volume of transmitted CSI data, thereby facilitating efficient classification through convolutional neural networks (CNNs). 

Although the proposed methodology is of considerable practical relevance and corresponds with the increasing demands for low-latency, bandwidth-efficient sensing within ubiquitous computing, various methodological and empirical elements necessitate enhancement prior to the work being deemed suitable for publication.

A principal advantage of the manuscript lies in its endeavor to connect the domains of compressed sensing and deep learning for the purpose of human activity recognition in indoor settings. The authors have developed a comprehensive pipeline that initiates with signal segmentation and preprocessing and culminates in the reconstruction and classification of compressed CSI data utilizing a trained CNN. The algorithm attains remarkable accuracy—exceeding 90%—while significantly diminishing transmission expenses. This suggests a pronounced potential for implementation in edge computing or Internet-of-Things (IoT) applications wherein communication overhead poses a substantial limitation.

Nevertheless, the asserted novelty of the KCS algorithm is inadequately distinguished from existing literature. Although the application of KSVD for dictionary learning is well-documented, particularly within signal reconstruction contexts, the manuscript does not distinctly articulate the methodological innovations that KCS introduces beyond current sparse representation techniques. A more rigorous analysis or benchmarking against cutting-edge alternatives within the dictionary learning realm (e.g., online dictionary learning, deep unfolding networks, etc.) is imperative. In the absence of this, the uniqueness of the approach risks being perceived as merely incremental.

The experimental design, although competent, exhibits insufficient scope to substantiate the generalizability of the findings. Data have been collected in merely two indoor settings involving a limited number of participants. This raises apprehensions regarding overfitting and constrains the assertion of extensive applicability. Furthermore, the manuscript fails to furnish adequate details concerning the sizes of the training, validation, and test splits, or whether any form of cross-validation was implemented. Additionally, there is a notable absence of error metrics beyond mere raw accuracy—such as confusion matrices, F1 scores, or standard deviations—which complicates the evaluation of the system's robustness under varying conditions. This consideration is particularly vital for applications that must contend with real-world noise and human variability.

The CNN classifier is succinctly outlined, yet its architecture lacks comprehensive analysis. A modular ablation study would be beneficial in evaluating the contribution of each preprocessing stage (e.g., WMA smoothing, segmentation threshold T) as well as the architectural choices made within the CNN. This assessment is crucial for isolating performance enhancements attributable to the KCS compression, distinct from downstream learning components.Moreover, the sensitivity of the pipeline to hyperparameter selections—such as the compression rate μ, sparsity level, and reconstruction thresholds—should be examined to ensure reproducibility and the feasibility of deployment.

The manuscript, although predominantly comprehensible, is impeded by grammatical inconsistencies and stylistic deficiencies that obscure its fundamental contributions. In particular, the abstract fails to articulate a compelling narrative or contextualize the research within a broader real-world framework. Numerous sentences throughout the document are constructed in passive or fragmented structures, and a comprehensive linguistic revision would significantly enhance both the readability and the scientific tone of the manuscript. Moreover, the conclusions presented are excessively optimistic and would benefit from a more nuanced discussion regarding the inherent limitations.

The ethical implications associated with the collection and utilization of Wi-Fi Channel State Information (CSI) data remain unaddressed. Although the technical facets of CSI-based recognition appear promising, they concurrently engender considerable privacy concerns, particularly in domestic or professional environments. There is an absence of discourse on data anonymization, user consent, or adherence to regulatory guidelines. These matters necessitate transparent discussion, especially in light of the escalating public and academic scrutiny surrounding surveillance technologies.

Reproducibility constitutes another significant area of concern. The manuscript does not provide access to datasets or source code, nor does it specify whether these resources will be made available to the public. This lack of accessibility undermines the possibility for replication and validation of the findings by independent scholars. A definitive statement regarding data and code availability is imperative.

In conclusion, the manuscript presents a relevant and potentially transformative technique for the transmission of compressed CSI data within activity recognition systems. Nevertheless, the contributions necessitate clearer articulation concerning their novelty, the experimental validation must undergo substantial expansion, and the overall presentation warrants enhancement. Ethical considerations and issues of reproducibility should be addressed to align with the standards of scientific transparency.

In light of these limitations, I recommend a major revision. The work demonstrates potential and may eventually warrant publication; however, significant improvements are requisite across the domains of methodology, evaluation, and exposition.

Comments on the Quality of English Language

The language could be improved.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The KCS algorithm, a novel compressed sensing (CS) method based on KSVD dictionary learning ids proposed, which enhances sparse representation and can achieve 90% accuracy with 90% data reduction. It addresses the critical challenge of high CSI data transmission overhead in Wi-Fi-based passive sensing systems, offering practical value for real-world deployment.  

Comments:
  1. What's the computational complexity, especially on resource limited devices?  
  2. The motion speeds (e.g., slow vs. fast walking) on compression performance could be explored.  
  3. How to guarantee data security for dictionary transmission? This should be clarified.  
  4. It is better to discuss the limitations of KCS.  
  5. Some typos should be corrected, e.g. c^k under equ.(10).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised manuscript is improved compared to the former version. My previous comments are well addressed, and the presentation is improved significantly. The composition pattern and some other ideas are well elaborated, making them clearer. Overall, I tend to accept this manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have provided additional experimental results that underscore the distinct aspects of the KCS algorithm compared to traditional sparse representation techniques. Although the suggestion to benchmark against more cutting-edge alternatives (e.g., online dictionary learning, deep unfolding networks) was not fully implemented, the authors argue convincingly that their experiments already highlight the unique adaptation of KCS for CSI data.

Data partitioning was also clarified and noted the use of 4-fold cross-validation. The autohrs acknowledged the limited scope of data collection and outlined plans for broader future data collection. These clarifications meaningfully address the concerns about generalizability and robustness.

Regarding the CNN Classifier and Ablation Studies the authors recognized the need for a more detailed evaluation of the CNN architecture and preprocessing steps. They commit to performing modular ablation studies in future work, which is an acceptable compromise given that the present study focuses primarily on the compression method.

The ethical concerns were also addressed in the revised version of the paper.

Finally, the authors clarified that while the dataset cannot be made publicly available due to confidentiality constraints, the source code will be accessible upon request (subject to partner requirements).

Overall, the authors have addressed the majority of the comments in a constructive and detailed manner. While some issues (especially advanced benchmarking of dictionary learning methods and more exhaustive ablation studies) are postponed to future work, the revisions considerably improve the manuscript’s clarity and rigor. Therefore I recommend acceptance.

Reviewer 3 Report

Comments and Suggestions for Authors

All coments have been resolved