To address the challenges of real-time detection and precise recognition of hazardous behaviors among construction site personnel, this study proposes a visual behavior recognition system tailored for smart construction sites. The system is based on the fusion of multi-source heterogeneous visual information and integrates YOLOv8 object detection, HRNet pose estimation, and ST-GCN temporal behavior classification to construct a four-layer architecture comprising "perception, processing, decision-making, and feedback". Experimental results demonstrate that the system achieves an average recognition accuracy of 89.4% across five categories of hazardous behaviors with an mAP@50 of 89.7%, thereby providing effective technical support for safety management in smart construction environments.

Xue Xiangyu. Research on Visual Behavior Recognition Technology for Smart Construction Sites. Academic Journal of Engineering and Technology Science (2026), Vol. 9, Issue 3: 165-169. https://doi.org/10.25236/AJETS.2026.090321.

[1] Tang Jiexin. Analysis on the Development of Safety Monitoring and Early Warning Systems for Smart Construction Sites [J]. Green Construction and Intelligent Building, 2026, (4):84-87.

[2] Du Qifeng. Research on a Data-driven Project Management Decision-making Model for "Smart Construction Sites" [J]. Brick - Tile, 2026, (4):100-102.

[3] Jia Haoming, Yan Fengwei, Wang Jie. Application of Smart Construction Site Technologies in Building Engineering Safety Management [J]. Sichuan Cement, 2026, (5):121-123.

[4] Chen Xianbin. Intelligent Recognition and Early Warning of Hazardous Behaviors on Construction Sites Based on Computer Vision [J]. Construction and Architecture, 2026, (4):116-118.

[5] Zhuang Tinglin, Ning Fuhua, Xie Qiaomin, et al. Intelligent Recognition and Management of Unsafe Behaviors among Construction Workers Based on Machine Vision [J]. Value Engineering, 2025, 44(18):48–51.