Leveraging semisupervised learning for domain adaptation: Enhancing safety at construction sites through long-tailed object detection

The advancement of deep learning has led to a growing demand and increase in research on computer vision-based construction site monitoring for improved safety and operational efficiency. These methods largely depend on supervised learning, requiring labeled data for optimal performance. However, when applied to new construction sites with varied environmental conditions, the effectiveness of these models is often compromised. Additionally, highly imbalanced object class distributions in the data sets, known as long-tailed objects, presents significant challenges during model training, considerably impacting performance. Recognizing this crucial gap in the field, this study proposes a novel approach to improve safety and operational efficiency at construction sites by leveraging a semisupervised learning approach for domain adaptation in long-tailed object detection. The method addresses the challenges of unbalanced class distribution and environmental variability in construction site monitoring, which often degrade the performance of computer vision models. By employing semisupervised learning, both labeled and unlabeled data are utilized in domain adaptation to unseen construction sites, considering both image-and object-level noise, thereby enhancing the model's adaptability to diverse working conditions. Based on the detection results, a risk scenario detection algorithm is also introduced for construction vehicles and workers. The efficacy of the proposed approach was validated through extensive experiments conducted on a comprehensive data set sourced from AIHub and CrowdHuman, in addition to actual self-labeled closed-circuit television (CCTV) data comprising 500 videos from construction sites' CCTV cameras. The evaluations revealed that the proposed method significantly outperforms conventional semisupervised learning by 9.76% on mean average precision for construction vehicle detection and by 3% for the worker detection model, paving the way for advanced construction site monitoring systems that ensure a safer and more efficient working environment.