Real-Time AI-Driven Fall Detection Method for Occupational Health and Safety
, Piotr Sowiński 1,2,*, Kajetan Rachwał 1,2, Karolina Bogacka 1,2, Anna Dąbrowska 3, Monika Kobus 3, Krzysztof Baszczyński 3, Małgorzata Okrasa 3, Witold Olczak 4, Piotr Dymarski 4, Ignacio Lacalle 5, Maria Ganzha 1,2 and Marcin Paprzycki 1Round 1
Reviewer 1 Report
This is an exciting paper focused on presenting a cutting-edge approach to fall detection using artificial intelligence approaches. The proposed method is claimed to potentially revolutionize workplace safety. Particularly, the article offers an AI-driven fall detection solution that can accurately detect the relevant fall scenarios. The system supports collecting real-time data from wearable devices, processing it in the edge-cloud continuum, including inference of an AI fall detection model, and logging the identified fall accidents. According to the authors, the system can improve worker safety and well-being in construction sites and similar environments.
The article is nicely written, although it contains a relatively brief analysis of related works, which is not particularly focused on comparing individual solutions but rather on an overview of approaches used in fall detection. However, I don't think it should matter. The solution is tested in laboratory conditions; it is valuable to focus the experiments on falls from a height, which, as the authors emphasize in several places, is not very often the subject of other authors' research. The use of an artificial figurine, structurally corresponding to a human, is fascinating.
The article also notes that deep learning models can perform well while operating only on raw acceleration data and skipping the feature handcrafting step. Still, they can lead to an increase in the computational cost and become a problem in the case when limited resources are available on the target wearable device.
The article lacks any more thorough justification for why the presented solution is better than others. Of course, exciting results are achieved with the help of experiments. However, in the discussion and conclusion, which are too short for my taste, there is still enough space to emphasize the merits and originality of the proposed solution.
Author Response
Please see attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments included in the attached pdf
Comments for author File: Comments.pdf
Author Response
Please see attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Overall: I appreciate the author's commitment to addressing all of my previous comments. I do not have any major concerns other than the minor comments listed below.
- Line 231 - “Due to the lesser risk of the tested activities, this study was able to collect data from both laboratory and outdoor environments, emphasizing that similar near-miss fall detection accuracy was achieved, regardless of the experimental environment.” Was data from outdoor environment collected for evaluation? Also, I am also not clear with what is being conveyed with the second part of this sentence “emphasizing that similar near-miss fall detection accuracy was achieved, regardless of the experimental environment.”
- Line 251 – What is meant by unique activity patterns, please give an example in brackets. Why is it important for this study?
- Figure 2 – Why were the tags connected to an RPI3 instead of the laptop directly? What is the role of the RPi here? Is it processing some information before relaying it to the laptop?
- Line 511 - “The proposed method uses a neural network composed of several recurrent layers and a dense output layer.” Please be precise and mention the number of layers and activation functions.
The paper is good in most parts but requires minor grammatical corrections in some sections. I would suggest the authors perform a full grammar and spell check for the next submission.
Author Response
Please see attachment.
Author Response File: Author Response.pdf
