Identifying unsafe behaviors of construction workers through an unsupervised multi-anomaly GAN approach

Unsafe behaviors of construction workers often lead to construction accidents, while traditional safety monitoring methods have low efficiency and timeliness. Although deep learning has been widely applied in detecting construction site unsafe behaviors, most studies still focus on supervised learning for detecting personal protective equipment use, with less emphasis on worker behaviors. This study developed an unsupervised learning model, Multi-Anomaly GAN, based on GAN with five pseudo anomaly synthesizers to detect unsafe behaviors of construction workers. First, the WeTeam22 dataset was built with tower crane edge scenarios to compensate for the lack of construction worker behavior datasets. After training on WeTeam22, Multi-Anomaly GAN achieved better results in AUC, EER, and F1 score compared to baseline methods. The model effectively identified unsafe behaviors during testing. The study proves the effectiveness of Multi-Anomaly GAN in detecting unsafe construction site behaviors, providing a novel and valid detection scheme for this problem.