Recurrent Neural Network-Based Hybrid Localization for Worker Tracking in an Offshore Environment

Round 1

Reviewer 1 Report

This paper describes a proposal for people tracking, from wireless signals and IMUs, in an offshore environment based on a Recurrent Neural Network. The paper is well organized and easy to read, and the topic is interesting. Furthermore, it presents several experimental results obtained in a real environment. However, there are some aspects which should be improved.

1. In the introduction it is indicated that the proposal achieved a 29% improvement (in terms of reduction in localization time tracking accuracy) in a deck, compared to existing state-of-the-art methods. However, there are not presented results these state-of-the-art methods. It will be interesting compare the proposal results against other state-of-the-art approaches, not only against the Viterbi method.
2. Figure 2 shows the structure of the proposed method. It includes the user device and the positioning server, but where is located this positioning server? Is it included in the user device or in another device? How is carried out the communication between modules?
3. I have some concerns related to the learning data collection and the ground truth estimation, since the step length can change for different people. Does the proposal work properly if the step length changes, or does it need to be retrained?
4. There are several constants in the algorithm, for example the w0 and w1 coefficients in equation 2. How were these values set? What is the effect of these values in the algorithm?
5. In the RNN training, it is used a mask because the length of the sequence can be variable Is this mask also used in test? What is the effect of modifying the number of input signals between train and test?
6. It is very interesting the experimental results in a real environment, but I have some questions about the presented results.
   1. There are presented the experimental results for 4 different decks in a boat but, is the algorithm capable of determinate the deck, or is it necessary to do the training and test for each deck separately?
   2. I guess that there is always the same person who takes the data with the mobile device, so the step length is similar in all the experiments. What happens if the step length changes?
   3. What are the effects of a signal missing? Does it require a new train? And what happen if there appears a new WiFI or BT signal?
   4. How is it obtained the ground truth? What is its accuracy?
   5. There are presented some results related to the computation time… what are the technical specifications of the device in which these time values are obtained?
7. The results are compared with those obtained with HMM-based Viterbi algorithm, but it is claimed that the proposal achieves a 29% improvement in a deck. For a fair comparison, it is necessary to evaluate other state-of-the-art algorithms in the same experimental environment.
8. Finally, a complete review of grammar and vocabulary is necessary.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposed RHLS, an indoor positioning system that estimates locations by fusing various embedded sensors of smartphones using an RNN. The experiments were conducted offshore at the Daewoo Shipbuilding & Marine Engineering shipyard. I feel that this paper is interesting and practical. My minor comments were listed below. First, I wonder the difference between the indoor building and an offshore environments? Does that reflect on the measurements? I would like to some discussion about this issue. Some competition paper were listed below for the authors reference. [1] Evaluating indoor positioning systems in a shopping mall: The lessons learned from the IPIN 2018 competition, IEEE ACCESS, 2019. [2] Off-line evaluation of mobile-centric indoor positioning systems: The experiences from the 2017 IPIN competition, Sensors, 2018. Next, it would be better to investigates the impact of the RNN parameters on the positioning errors, such as the network structures and the types/directions of LSTM used in the experiments.   

Author Response

please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The author has addressed most of the reviewer's comments, improving the content and style of the paper. I agree with this version of the manuscript.