A Study of Detection Systems for Safety Guardrails at Construction Sites: Innovative Applications of Image Processing and Perspective Correction

In response to the increasing demand for intelligent safety management in construction engineering, this study develops an automated detection system for construction-site safety guardrails that integrates advanced image processing and perspective correction with deep learning techniques. Traditional manual inspections are labor-intensive, error-prone, and lack real-time responsiveness, while most existing deep learning approaches (e.g., YOLO-based models) mainly rely on visual features and often suffer from perspective distortion, camera angle variation, and complex background interference. To overcome these limitations, we propose a dual-stage detection framework that incorporates YUV color space transformation, contrast-limited adaptive histogram equalization (CLAHE), and edge-based perspective correction to stabilize detection across diverse site conditions. A vertical railing detection model is first applied, followed by geometric correction, and then a horizontal railing detection model refines the structural consistency of the identified guardrails. Experimental validation demonstrates that the proposed method outperforms conventional YOLOv4-based approaches by improving the overall accuracy to 93.8% and F1 score to 87.5%, especially under challenging illumination and perspective scenarios. Compared with prior studies that focus primarily on object detection, this research emphasizes the integration of geometric constraints and perspective correction, thereby enhancing robustness in steel-structure construction environments. The system provides a scalable and intelligent solution for automated safety inspections and offers practical implications for innovative construction. Future work will explore integration with building information modeling (BIM) platforms, LiDAR-assisted depth sensing, and semisupervised learning to improve adaptability further and reduce deployment costs in real engineering contexts. The primary contribution of this study lies in its hybrid framework that combines deep learning, image preprocessing, and geometric perspective correction, providing a more robust and practical solution for automated safety guardrail detection in complex steel-structure construction environments.