A Remote-Vision-Based Safety Helmet and Harness Monitoring System Based on Attribute Knowledge Modeling

Round 1

Reviewer 1 Report

By applying the transformer-based end-to-end network with a self-attention mechanism in attribute knowledge modeling, the authors present an advanced safety monitoring system. The authors argue the approach has practical application and can be an effective and efficient safety recognition system.

 

The authors conducted a solid literature review, designed their experiments carefully, and validated the model. However, below are comments/suggestions for further improving the manuscript:

 

1. Consider revising the title so it properly reflects the project scope and outcomes. 'Safety-monitoring system' can be misleading when the model only detects PPE requirements for fall protection.

 

2. The authors state that the object detection method for safety monitoring cannot meet the demands of practical application scenarios (lines 56 to 58). Please provide the basis for this statement. If this statement is to merely justify the presented study, revise the statement by pointing out shortfalls of the object detection approach rather than claiming they're not going to work. The authors should refer to more recent relevant studies, DOI: 10.1109/ACCESS.2021.3135662, 10.3390/drones6090222, and 10.1016/j.jsr.2022.09.011.

 

3. In conclusion, the authors claim their safety-monitoring system to be novel and practical by identifying whether workers wear safety helmets and harnesses. However, workers under fall hazards merely wearing helmets and harnesses are not considered properly protected under the personal fall arrest system if no lifeline connected to the harness and secured to an anchor point. The authors should add this crucial component of the PFAS within their attribute recognition model in processing images.

 

4. While the authors claim the practicalness of their safety-monitoring system, how would video stream data be collected if workers are exposed to fall hazards where the video surveillance camera network is not reachable (e.g., high-rise building construction)? What about using a computer-vison-based drone?

 

 

5. Letters in Figures 10 & 11 are too small to read.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presents a remote vision-based safety-monitoring system that transfers the traditional object detection problem to the semantic attribute recognition problem for identifying the wearing states of safety helmets and harnesses. The attribute recognition method adopts an attribute knowledge modeling network based on the transformer architecture to explore the relationship between attributes and image features. To address the lack of the safety-monitoring data set, an open-site safety-monitoring data set on construction scenes was contributed for the evaluation of safety-monitoring systems in unconstrained environments. A series of experiments were conducted, including object detection, object tracking and safety attribute recognition. My comments were given as follows.

1.     The description of Figure 4 is not clear to me and should be described more clearly in the text.

2.     The font size of the software part in Figure 1 is a little bit small, which may make the reader hard to read.

3.     Subsection 3.3.2 mentions the position encoding P, but some introduction on how to get P should be given.

4.     Equation (4) should be rewritten as the variable l is confusing.

5.     YOLOv5 was mentioned in Page 5 and Page 13, but the expressions were kind of different.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Figure 2, uses UML

Figure 3, improve readibility

Figure 4, use any form of UML

Raw 263, Github cited improperly, indicate author of the dataset

Literature review must be improved with international references, and recent 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf