TY - GEN
T1 - Detecting personal protective equipment (PPE) utilising YOLOv8 in a federated learning environment
AU - Makris, Ioannis
AU - Lytos, Anastasios
AU - Kyranou, Konstantinos
AU - Argyriou, Vasileios
AU - Lagkas, Thomas
AU - Kollias, Konstantinos Filippos
AU - Fragoulis, George F.
AU - Sarigianndis, Panagiotis
PY - 2024/1
Y1 - 2024/1
N2 - In the construction industry, where worker safety is a paramount concern, the effective use of PPE is crucial. This study delves into an innovative approach by exploring the integration of Federated Learning (FL) with the YOLOv8 architecture for enhanced PPE object detection. A specialized construction safety dataset was utilized to conduct the corresponding experiments, comparing FL against conventional local training methodologies. The results have clearly shown the superiority of FL in terms of accuracy improvement of PPE object detection which opens a lot of paths for creative solutions leading to the progress of safety performance at construction sites, at the same time, while observing the strictest privacy and security issues for personal data that is a primary consideration in our data-driven era. The proposed solution achieved an average mAP of 79.52%, and an average Recall of 78.49%, proving its effectiveness over the centralized training.
AB - In the construction industry, where worker safety is a paramount concern, the effective use of PPE is crucial. This study delves into an innovative approach by exploring the integration of Federated Learning (FL) with the YOLOv8 architecture for enhanced PPE object detection. A specialized construction safety dataset was utilized to conduct the corresponding experiments, comparing FL against conventional local training methodologies. The results have clearly shown the superiority of FL in terms of accuracy improvement of PPE object detection which opens a lot of paths for creative solutions leading to the progress of safety performance at construction sites, at the same time, while observing the strictest privacy and security issues for personal data that is a primary consideration in our data-driven era. The proposed solution achieved an average mAP of 79.52%, and an average Recall of 78.49%, proving its effectiveness over the centralized training.
KW - Computer science and informatics
U2 - 10.1063/5.0236071
DO - 10.1063/5.0236071
M3 - Conference contribution
VL - 3220
T3 - AIP Conference Proceedings (2024)
BT - ETLTC2024 international conference series on ICT
PB - AIP Publishing
T2 - The 6th ETLTC International Conference on ICT Integration in Technical Education
Y2 - 23 January 2024 through 26 January 2024
ER -