Real-Time Detection of Personal Protective Equipment Violations for Construction Workers Using Semisupervised Learning and Video Clips

作者：Q. H. Chen, Danbing Long, Siqi Wang, Qirong Chen, Beifei Yuan · 发表于：Journal of Construction Engineering and Management · 年份：2024 · DOI：10.1061/jcemd4.coeng-15310 · 被引用次数：18 · 研究领域：Occupational Health and Safety Research、Risk and Safety Analysis、Infrastructure Maintenance and Monitoring

The collection of personal protective equipment (PPE) violation data is crucial for assessing safety risks in behavior-based safety (BBS) management on construction sites. Owing to the limitations of manual inspection methods, many studies have employed computer vision–based methods for PPE violation detection. However, limitations exist in implementing this in actual construction projects due to the costs associated with the acquisition and labeling of a large number of images, and the accuracy and efficiency of recording PPE violation data. Therefore, this study introduces semisupervised object detection (SS-OD) and data augmentation for training PPE detection models to reduce the use of labeled data without reducing the performance, and proposes a framework for recording the PPE violation data by extracting PPE violation videos from real-time surveillance footage by considering the role of the worker and their position, thereby enhancing practical construction management and serving as input for BBS-based safety risk assessments on construction sites. The results show that (1) the labeled data demand for training PPE detection models can be reduced through semisupervised learning, image augmentation, and transfer learning, without reducing the performance of PPE detection, thereby reducing the cost of application on actual construction sites; (2) SS-OD methods are better equipped to handle changes in construction scenarios by making full use of unlabeled data, and thus are...