DEVELOPMENT OF AN INTEGRATED COMPUTER VISION FRAMEWORK FOR REAL-TIME CONSTRUCTION WORKER SAFETY MONITORING

ABSTRACT: The construction industry consistently ranks among the most hazardous occupational sectors globally, with over 90% of workplace accidents attributable to unsafe worker behaviors and hazardous working conditions. Traditional safety management approaches relying on manual inspections are inherently limited by their labor-intensive nature, inability to provide continuous monitoring, and reactive rather than proactive orientation. This review paper synthesizes findings from peer-reviewed literature published primarily between 2017, examining the evolution and current state of computer vision technologies for construction safety monitoring. The review identifies that deep learning-enabled visual computing applications in construction are dominated by object detection (33%), construction safety (28%), damage detection (22%), construction quality (9%), and productivity analysis (8%). The YOLO family of algorithms has emerged as the most widely adopted architecture for real-time detection, with YOLOv8 achieving the best balance of accuracy (mAP 0.969–0.971) and computational efficiency. Pose estimation techniques, particularly those integrating CBAM attention and deformable convolutions, enable ergonomic risk assessment by identifying awkward postures associated with work-related musculoskeletal disorders, which account for over 20% of nonfatal injuries in construction. Identity association remains a critical challenge, with multimodal integration of computer vision and RFID/UWB technologies demonstrating significant improvements—up to 20.3% during severe occlusion and 18.8% during worker overlap conditions. Integrated frameworks combining detection, pose estimation, behavior recognition, and identity tracking achieve overall F1-scores of 0.904 across all behavior categories, with PPE violation detection being most reliable (F1-score: 0.950) and ergonomic violation detection most challenging (F1-score: 0.839). Real-time processing at 30 FPS on high-performance GPUs and 15 FPS on edge devices demonstrates practical deployment feasibility. Key research gaps include integrated multi-task monitoring, behavior-identity association under occlusion, complex scene robustness, and multimodal integration strategies. The review concludes that integrated computer vision frameworks offer a promising pathway toward comprehensive, automated construction safety monitoring that can complement traditional safety management approaches. Keywords: Construction safety, computer vision, deep learning, object detection, pose estimation, behavior recognition, identity tracking, YOLO, real-time monitoring, PPE detection, ergonomic risk assessment, multimodal integration