Citation

Li J, Zhao X, Zhou G, Zhang M. Safety Sci. 2022; 150: e105689.

Copyright

(Copyright © 2022, Elsevier Publishing)

DOI

10.1016/j.ssci.2022.105689

PMID

unavailable

Abstract

Falling from height (FFH) and object strike (OS) accidents frequently occur at construction sites, threatening personnel safety and affecting the construction quality. In the hierarchy of controls, personal protective equipment (PPE) is the most easily achievable measure. However, the presence of hazards and the differentiation in subjective protection awareness make PPE the least effective control measure. Therefore, it is necessary to promote proper and standardized use of PPE to meet the requirements from the perspective of administrative control. Taking two behaviors that can lead to OS and FFH accidents as a research case--loosening the hardhat and not using the safety harness's hook--this study proposes a deep-learning-based inspection method. First, we established a detection model for hardhats and hooks based on You Only Look Once v5. Thereafter, the object-detection model and Openpose algorithm were applied to recognize 1200 video clips containing three unsafe behaviors and one safe behavior and generate 1200 data files that vary in a time series. Finally, a one-dimensional convolutional neural network (1D-CNN) model was trained with the data of 600 videos, and the model was used to test the data of the other 600 videos. The accuracy attained was 0.9467 in the experimental scenario. Using the proposed method, the improper use of PPE can be determined without affecting the normal behavior of individuals, which can improve the efficiency of safety management.

Language: en

Keywords

Convolutional neural network; Deep learning; Fall prevention; Object detection; Openpose; Personnel protective equipment