Reviewer 3:

Comment 1

There are grammar issues, e.g., mistakes of singular and plural forms; please correct them in the revised manuscript.

Answer 1

I am grateful for this crucial comment, and I have revised the manuscript based on the comments. The grammatical and spelling errors have been checked carefully.

Comment 2

The image resolution of figure 1 is too low

Answer 2

Thank you very much. Figure 1 is replaced by a new one.

Comment 3

The methods listed in Table 1 should be described clearly. E.g,, how are features extracted from video clips? Which layer are the features extracted from? What are the feature dimensions? How are the CNN models used for classification? Do you replace the last classification layer and re-train/fine-tune the CNN? In general, there are important details missing in the method and experiment sections

Answer 3

Thank you very much for the decisive comment. The feature dimension is employed using the matric of Action feature (Eq.4). In addition, I have added detailed information about CNN model used for classification.

Quotation from (Page 9, Section 4.3.2 (

Reviewer 1:

Comment 1

Moderate English changes required

Answer 1

I am grateful for this crucial comment, and I have revised the manuscript based on the comments. The grammatical and spelling errors have been checked carefully.

Comment 2

The authors need to more clearly introduce the technical novelty

Answer 2

I thank the reviewer for the constructive comment. In the revised version of the manuscript, every attempt has been made to make the technical novelty clearer.

Quotation from (Pages 2 and 9, Sections 1 and 4.3.2(

Comment 3

It will be interesting if authors can compare the method with RGB video based solutions. Using IoT leading to better or worse results?

Answer 3

Thank you very much for the important comment. The system was implemented using the UR fall detection dataset and promising results were obtained, in which using IoT led to better results. In fact, this point is very important, since I will work on developing it in the future. I have elaborated on this matter in the updated version of the manuscript.

Quotation from (Pages 13, Sections 6(

Comment 4

Authors also need to discuss the disadvantages of using IoT for human action analysis. For example, can this be used for long-distance action recognition? Say, action recognition from UAV. Author are suggested to discuss this.

Answer 4

Thank you very much for the constructive comment. In fact, gathering real-world dataset from IoT camera devices is intended as part of the future work. This dataset can contain more challenging images such as emergency incidents images that contain crowds and other emergency incidents images captured from long distances (e.g., images collected by drones or Unmanned Aerial Vehicles (UAVs)), and compare current results with the more challenging ones. I have discussed these ideas in the updated version of the manuscript.

Quotation from (Pages 13, Sections 6(

Reviewer 3:

Comment 1

English language and style are fine/minor spell check required

Answer 1

I am grateful for this crucial comment, and I have revised the manuscript based on the comments. The grammatical and spelling errors have been checked carefully.

Comment 2

While the author states what they feel the contribution of the manuscript are, the only potentially novel research contribution is the fall classification algorithm. The architecture for connecting the fall classification to emergency responders is an engineering implementation problem, not a research problem. It is also not clear if the architecture was implemented or tested

Answer 2

Thank you very much for this comment.  In fact, the system was implemented using the UR fall detection dataset and promising results were obtained. And the proposed work represents an important research point according to the contribution mentioned on Page 2, Section 1. In addition, more details about the implementation have been added on Pages 10-11, Section 5.

Quotation from (Pages 2, 10 - 11, Sections 1 and 5 (.

Comment 3

The manuscript can be considerable shortened by focusing on what is potential novel about the work that is the fall identification algorithm. The emergency response architecture should be removed from the manuscript.

Answer 3

Thank you very much for the constructive comment.  In fact, this point is very important, since more details on the implementation of the emergency response architecture have been added to the updated manuscript, and therefore I see the inevitability of its existence.

Quotation from (Pages 10 - 11, Section 5 (.

Comment 4

More detail is needed on the architecture of the CNN, such as number and types of layers, and training parameters. A diagram would be appropriate for the whole CNN/SVM.

Answer 4

In addition, more detailed information about the CNN model used for classification is added in the updated manuscript ;

Quotation from (Page 9, Section 4.3.2 (

Comment 5

The labels on the confusion matrices do not make sense. Which cells are the true positives, and which are the true negative?

Answer 5

Many thanks for your comment, In Figure 5” Confusion matrix for EfficientNetB0, MobileNetV2, ResNet50, Xception, method 1 and method 2” , A confusion matrix is a grid of information that shows the number of True Positives [TP] cell22, False Positives [FP] cell 21, True Negatives [TN] cell11, and False Negatives [FN] cell12.

Reviewer 3:

Comment 1

The author clarified the contribution of the architecture in the current version of the manuscript, and that the architecture was implemented. The author now provides the architecture of their system.

Answer 1

Thank you for this comment. I am glad that the previous comment has been addressed.

Comment 2

The author’s response to this reviewer indicates which cells correspond to TP (cell22), FP (cell 21), TN (cell11), and FN (cell12). However, the labelling on the confusion matrices in Figure 6 still does not convey that information. Something like the following would help the reader.

Answer 2

Thank you very much for your observation, and for this reason, Figure 6 has been re-clarified to remove any confusion and ambiguity

Comment 3

The author did not address this comment from my previous review: While the fall classification algorithm is compared to several other methods, it is not clear whether these methods were previously applied to the fall database that the author uses. Since the database is public, there must be other authors who have evaluated fall prediction on the database. The author should compare to these methods.

Answer 3

Thank you very much for the decisive comment. Yes, this is true, the comparison presented in this work is between different CNN architectures including EfficientNetB0, MobileNetV2, ResNet50, and Xception. Additionally, we used two different methods including method 1 where the SVM classifier was used to identify the features only and method 2 where CNN and SVM are combined for the whole process. To the best of my knowledge, I am the first to use such deep learning techniques on the UR fall dataset. Therefore, I believe other authors who are planning to use deep learning techniques on this dataset can compare with my work which will make the comparison fair.