CNN-Based Fall Detection Strategy with Edge Computing Scheduling in Smart Cities

Round 1

Reviewer 1 Report

The authors raise a very current and very interesting topic, especially due to the aging of the society.
Literature review - state of art - well conducted. However, I have the impression that the authors are not microelectronics. What do I conclude from. In their work, they postulate the use of smartphones as a key medium for obtaining data based on sensors embedded in the smartphone, e.g. accelerometers. Unfortunately, the sensors used in the phones have quite poor parameters in terms of response speed and sensitivity. However, this is not the only problem that I found in the authors' study. Location of the telephone. Where you have to wear it for a real model. Every change, e.g. in a pocket, or putting the phone into a purse or taking it in a hand, causes changes in the posturographic model. Additional complexity of the model appears. Exaggerating: A phone at knee level will not receive all the signals from your entire body that could determine an impending fall, just as if it is placed in your hand. That's why I missed it here. The second part of the article on the use of neural networks is fine. It works for this model. The question, however, is to what extent the results obtained, with the problems I mentioned, will actually be used in the actual detection of falls.

Author Response

Thanks for your appreciation on accepting our paper. At first, as the reviewer said, our major is computer science rather than microelectronics. We are following the fall detection problem based on the computer technology including but not limited to the field of microelectronics. Secondly, although the sensors used in the phones have quite poor parameters in terms of response speed and sensitivity, we in this paper use SDN to grasp the global network view in order to guarantee the on-line data availability. In addition, for these parameters, the CNN can be used for this well. Thirdly, in our experiments, the smartphone is taken in a hand, and the experimental results only consider this situation. In fact, the major contributions concentrate on SDN, edge computing and CNN training, that is to say, we try to leverage the emerging hot technologies to address fall detection problem. Finally, CNN is used to train the obtained history data and to conduct the fall detection, in which the convergence must be completed so as to obtain the subsequent results.

Reviewer 2 Report

The manuscript by Pan et al., titled, 'CNN-based Fall Detection Strategy with Edge Computing Scheduling in Smart City' describes their proposed method to detect fall from smartphone data using Convolutional Neural Network (CNN) which the authors claim can work well with edge computing and collects data through Software-Defined Networking (SDN).

Though the topic is of wider interest for a broad audience due to the increasing number of elderly people in the world, the manuscript has several serious flaws, as listed below, due to which it can not be published in its current form.

• The paper requires major rewriting. The English of the paper must be improved. In its current form, the writing is quite clumsy and many of the sentences don't make sense. It must be reviewed by Language editors.
• Fig. 1 shows the overall system framework. However, the hypothetical architecture appears to be flawed. It is not clear where CNNs are being employed and how SDN is enabling data generation. Also, how edge computing is performing the early data analysis is not clear. This requires major rewriting.
• The methodological contribution is not clear. What is the novelty of the work? Is it the overall pipeline? Is it the data acquisition or the signal analysis? This has to be very clearly written in the text.
• Also, the feature extraction method requires a more elaborated explanation.
• It is not clear how the methods to which the current method was compared were working? Also, the authors need to implement the methods for comparison rather than getting some published value on some predefined dataset.
• There must be a detailed discussion of the pros and cons of the current pipeline.
• The overall contribution of the work is poor. The novelty of the paper is limited and it lacks clarity in its writing and organisation. 

Author Response

Q1: The paper requires major rewriting. The English of the paper must be improved. In its current form, the writing is quite clumsy and many of the sentences don't make sense. It must be reviewed by Language editors.

R1: In the revised version, we have carefully modified and improved the writing, and they have been highlighted in Red color.

 

Q2: Fig. 1 shows the overall system framework. However, the hypothetical architecture appears to be flawed. It is not clear where CNNs are being employed and how SDN is enabling data generation. Also, how edge computing is performing the early data analysis is not clear. This requires major rewriting.

R2: Thanks for this comment. At first, CNN is employed both edge computing server and smartphone. Please see lines 178-179 and 186-187 for details. Then, the data collection of SDN is

realized by the OpenFlow switch. Please see lines 196-198 for details. Well, the control signals of SDN are generated automatically. Finally, the edge computing does not perform the early data analysis and it only performs the scheduling strategy. Please see lines 184-186 for details.

 

Q3: The methodological contribution is not clear. What is the novelty of the work? Is it the overall pipeline? Is it the data acquisition or the signal analysis? This has to be very clearly written in the text.

R3: Many thanks for this comment. In terms of the methodological contributions, this paper shows three aspects. (i) The global network view ability of SDN is used to collect the generated data from smartphone. (ii) The edge computing is exploited to put some computation tasks at the edge server by the scheduling technique. (iii) CNN is equipped with both edge server and smartphone to train the related data and give the guidance of fall detection. Please see lines 64-71 for details.

 

Q4: The feature extraction method requires a more elaborated explanation.

R4: We understand the worry of the reviewer. The feature extraction method in this paper is the principal component analysis which is a well-known method. Therefore, in order to save space, the readers can refer to reference [50] for details. Regarding this, we have stated that “In this paper, for each axis, the five features, i.e., minimum, mean, maximum, kurtosis and skewness are extracted, and the principal component analysis is employed for this. The concrete method can be found in [50].” We sincerely hope the reviewer could accept this.

 

Q5: It is not clear how the methods to which the current method was compared were working? Also, the authors need to implement the methods for comparison rather than getting some published value on some predefined dataset.

R5: Thanks for this comment. This comment is very nice. In fact, we very want to implement the methods for comparison. However, the source codes are difficult to be obtained. Therefore, in this paper, we only can refer to the predefined datasets for comparison, which is also very significant. We sincerely hope the reviewer could accept this.

 

Q6: There must be a detailed discussion of the pros and cons of the current pipeline.

R6: In the revised version, we have enhanced the pros and cons of the current pipeline in Section 7, as follows: “Pros: two kinds of simulation experiments are made. At first, the classification performance is verified via evaluating recall ratio, precision ratio and F1 value. Then, the whole performance is evaluated by testing accurate rate, transmission delay and stability. The experimental results show that EdCNN outperforms two baselines.

Cons: However, as a novel method based on SDN, edge computing and CNN, the proposed EdCNN also has some limitations. At first, we do not consider the application types, that is, the data division is not completed in the fine-grained way. Then, although CNN computing and edge computing decrease the communication delay, they introduce the computation overhead. Finally, the experiment environment is stalled at the simulation platform, irrespective of the real data collection from some persons. In future, we plan to enhance the performance of EdCNN from two aspects. On one hand, we improve EdCNN around the above mentioned three limitations. On the other hand, we improve BSA and reach much faster convergence.”

 

Q7: The overall contribution of the work is poor. The novelty of the paper is limited and it lacks clarity in its writing and organisation.

R7: Thanks for this comment. This paper uses SDN, edge computing and CNN to help fall detection. The whole contributions are very novel, as follows: “This paper proposes CNN-based fall detection strategy with edge computing, called EdCNN. In particular, the global network view ability of SDN is used to collect the generated data from smartphone. The contributions of this paper are recognized as follows. (i) The edge computing is exploited to put some computation tasks at the edge server by the scheduling technique. (ii) CNN is equipped with both edge server and smartphone to train the related data and give the guidance of fall detection. (iii) The simulation experiments are made by testing accurate rate, transmission delay and stability to demonstrate the efficiency.”

Reviewer 3 Report

The authors investigated CNN architecture for fall detection in edge computing scheduling in smart city.

The proposed approach is interesting but there are some points that the authors should consider.

The main issues about this paper concerns the methodology section in which is not clear how their methodology differs from the traditional one. Furthermore, the authors should be revised the introduction for underlying the novelties of the proosed approach. 

Furthermore, the authors should provide more details and discussion about the obtained results and used technologies.

I suggest to analyze also more recent approaches about the examined topics. In particular, I suggest to underline the relevance of CNN in different domains and security issues by further investigating these approaches, :

1) An emotional recommender system for music. IEEE Intelligent Systems.

2) Recognizing unexplained behavior in network traffic. In Network Science and Cybersecurity (pp. 39-62). Springer, New York, NY.

Finally, I suggest to perform a linguistic revision.

Author Response

Q1: The main issues about this paper concerns the methodology section in which is not clear how their methodology differs from the traditional one.

R1: Thanks for this comment. In fact, we have summarized the difference between our research and the traditional methods in last paragraph of Section 2, as follows: “According the above reviewed literatures, it is observed that the new emerging techniques have being used to help fall detection. In spite of this, the current research achievements always have some limitations which need to be improved and enhanced. For example, the data collection cannot reach the real-time level; the adopted detection methods cannot satisfy the enough stability; and the computation overhead of collection device is very high. Given this, this paper will exploit the emerging techniques (e.g., CNN and edge computing) and networking paradigms (e.g., SDN) to further study fall detection.” In other words, SDN, edge computing and CNN are novel elements in this paper.

 

Q2: The authors should be revised the introduction for underlying the novelties of the proposed approach.

R2: Thanks for this comment. The novelty of this paper has been summarized in lines 64-72. Especially, we have emphasized that “To the best of our knowledge, this paper is the first to do fall detection with SDN, edge computing and CNN.”

 

Q3: The authors should provide more details and discussion about the obtained results and used technologies.

R3: Thanks for your comment. The obtained results are discussed in lines 431-436. The used technologies are discussed in lines 423-430.

 

Q4: I suggest to analyze also more recent approaches about the examined topics. In particular, I suggest to underline the relevance of CNN in different domains and security issues by further investigating these approaches

1) An emotional recommender system for music. IEEE Intelligent Systems.

2) Recognizing unexplained behavior in network traffic. In Network Science and Cybersecurity (pp. 39-62). Springer, New York, NY.

R4: Thanks for this comment. At first, we have analyzed 23 papers, from [13] to [41], which have shown the latest research achievements. Then, we cite the two papers as the references on the relevance of CNN in different domains and security issues in last paragraph of Section 7, as follows: “In addition, we will study the relevance of CNN in different domains and security issues, like [53] and [54].”

 

Q5: I suggest to perform a linguistic revision.

R5: We have tried our best to perform a linguistic revision. All modifications have been highlighted in Red color.

Round 2

Reviewer 2 Report

The authors have handled all comments.

Reviewer 3 Report

The authors have addressed all my concerns.