Real-Time Detection of Unsafe Worker Behaviors via Adaptive Vision Transformers in Construction Sites

Abstract

Unsafe behavior of workers is a leading cause of construction accidents. However, existing monitoring systems remain limited by low efficiency and poor adaptability to dynamic on-site environments. This study proposes an adaptive dual-stream vision framework that integrates Dynamic Adaptive Image Enhancement (DAIE) and a Lightweight Real-Time Behavior Network (LR-BehaviorNet) to improve the accuracy and responsiveness of unsafe behavior detection. The DAIE module dynamically adjusts brightness, contrast, and sharpness according to scene conditions, ensuring visual clarity under varying lighting and weather. LR-BehaviorNet combines efficient convolutional blocks with Transformer-based temporal modeling to identify critical actions from both enhanced and raw image streams. Additionally, an adaptive thresholding mechanism fine-tunes detection sensitivity under complex visual interference. Experiments using open-source construction datasets demonstrate that the proposed framework outperforms conventional models—including Faster R-CNN, YOLO, and Mask R-CNN—in precision, recall, and F1-score, achieving 93.2%, 91.4%, and 92.3%, respectively. These results validate the robustness of the proposed method for real-time safety supervision and its potential integration with intelligent construction management platforms. Overall, the framework offers a scalable and efficient solution for automated safety monitoring, advancing the digital transformation of construction safety management.