Round 1

Reviewer 1 Report

review of IQ Data based WiFi Signal Classification Algorithm using the Choi-Williams and Margenau-Hill-Spectrogram Features: A Case in Human Activity Recognition

Review

This paper describes a fascinating way to learn human activities from a wifi signal. The ML learned model has a great accuracy of .95 and much potential.

I missed a comparison to state of the art for this problem in the conclusions. Is it an improvement? And if so, what caused the improvement?

suggestions

Some sentences are somewhat harder to follow.

A video showing the exact experimental setup with the human in action would be helpful.

More like Figure 1 would be exptremly helpful, giving the reader an indication that this problem is indeed solvable by a classifier. Can we see these for all 4 classes?

Why not do t-SNE to 2 dimensions and plot the 12K samples and color them with the true class. So we get a good quick idea of the difficulty of th eproblem.

needed improvements

• classification rate on page 6 is not defined (because E is not). In any case, with 4 classes, accuracy is the correct measure here for figure 4.
• accuracy makes no sense in figure 5. use F1 measure here.
• The experimental setup is not complete. W only have this sentence The holdout cross-validation partition (p = 0.3) is used via selecting 70% samples for 175 learning features and the remaining 30% for testing.
• table 2, typo for arms rotating accuracy. Also here F1 is much better to use. And why not precision of the true class?
• The description of the classes is not sufficient. Especially for the "empty signal". What is that? A person in rest? No person at all.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Review of manuscript IQ Data based WiFi Signal Classification Algorithm using the Choi-Williams and Margenau-Hill-Spectrogram Features: A Case in Human Activity Recognition

For MDPI Journal Electronics

The authors have developed a machine-learning (ML) scheme to classify a few Activities of Daily Life (ADL) using, as inputs to a random decision forest ensemble, statistical moments and principal components of images generated from two time-frequency distributions, the Margenau-Hill and the Choi-Williams. The time-frequency distribution images are spectrograms of in-phase I(t) and quadrature Q(t) components of IEEE 802.11 (Wi-Fi) signals in the 2.4 GHz band, sampled using a Tektronix RSA 306B spectrum/signal analyzer. The authors used the 2.4GHz band due to this band’s better solid object penetration characteristics than the 5 GHz band. The author’s scientific contribution is a novel classification technique to identify human subject ADLs using only Wi-Fi signals, which is useful since Wi-Fi infrastructure is commonly installed in institutions.

Questions for authors:

1. Did the authors evaluate classification accuracy using only one of the two time-frequency distributions, Margenau-Hill or Choi-Williams, to see if both are required to yield good results, or if one of the two is substantially superior to the other?
2. Where is the Wi-Fi access point (AP) with respect to the human subject? In the same room or in another room separated by a concrete wall?
3. My main concern with the author’s methodology is multiclass detection ability. The authors have measured classification performance in an environment with only one human subject. But a more realistic scenario would involve multiple human subjects in a room. How accurate is the author’s methodology in simultaneously detecting and classifying the ADLs of multiple persons, conceivably engaging in different ADLs (e.g., one person marching, another skipping rope, another idle, etc.)?
4. The authors state they use the 2.4 GHz band. However, the authors should be more specific and specify which of the 12 standardized channels they use in the 2.4 GHz band, and which specific channel(s) yielded better classification results. (e.g., channel 1 has center frequency 2.412 GHz, channel 2 has center frequency 2.417 GHz, etc.).
5. In a typical real-life scenario, the 2.4 GHz band will be shared among mobile device users in doors, potentially moving about and changing location. How does the presence of mobile device users that share a channel affect ADL classification? How does the presence of multiple mobile devices appear in the two spectrogram images, and can mobile devices be identified and filtered out from the images?
6. Marching-in-place and jump roping are somewhat similar exercises (both involve raising and lowering of the legs). Did the authors investigate other activities of daily life, such as walking, sitting down, rising from a chair, opening a door, etc.?
7. We see in Figure 8 that the contribution of using principal components as features yields a somewhat minor improvement in classification rate. How many principal components derived from SVD were selected for use as features? The authors should plot the singular values from slices on a log scale to show how much energy resides in each principal component and to show how much variance in the data can be captured in a few of the principal components.

Recommended minor corrections:

1. The flow chart in Figure 2 is symmetric about the y-axis. The authors can simplify Figure 2 and show one half of the flowchart and explain their method uses the same flow (or algorithm) for both time-frequency distributions.
2. On line 145 the authors should explain what they mean by the phrase “rich information of the unit”. What is “unit”? A subset of a spectrogram image?
3. Instead of using the phrase “marching on the spot”, write “marching-in-place exercise”.
4. What foot pattern of rope skipping was performed? There are many patterns such as basic jump, alternate foot jump, boxer step, high knees (similar to marching-in-place), etc.
5. On line 163 the authors write “empty” when they should write “idle” or “motionless”.
6. In table 1 there is a spelling error: the authors write “sport” when they intended to write “spot”.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper applies WiFi IQ data and its time frequency images to recognise human activities. Overall, the paper was well-written, the experimental results are solid and well-presented. Below are some minor comments regarding the presentations.

1) The "Related work" section is completely missing. It's mixed within the introduction section. The authors should clearly separate it.

2) There must be a "Paper's contributions" sub-section (within the Introduction) to clearly outline, using bullet points, the main contributions of the paper.

3) Texts in Fig. 3 are too small to see.

4) The numbers in first row of Table 1 should be centered.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf