YOLO Based Abnormal Behavior Recognition Methods for Construction Site Personnel

Abstract

The construction industry is a high-risk sector where effective safety monitoring is essential for reducing accidents and improving management efficiency. This paper presents a comprehensive review of anomalous behavior recognition methods for construction site personnel, focusing on the You Only Look Once (YOLO)-based algorithms for helmet violation detection, facial recognition for identity association, and temporal modeling for unrecovered fall detection. By systematically comparing representative studies, this work evaluates detection accuracy, real-time performance, and false alarm control of different approaches, while analyzing their robustness in complex environments. Based on recent research developments, the paper further summarizes key challenges in multi-task integration and temporal modeling, and discusses future directions for improving intelligent construction safety monitoring systems. Specifically, this study synthesizes cross-dataset evaluation protocols, proposes a unified benchmarking framework, quantifies latency-accuracy trade-offs, and highlights deployment guidelines for edge devices, enabling reproducible comparison and practical adoption in modern large-scale construction safety monitoring applications.