VISION-BASED AUTOMATIC MONITORING OF WORKERS FROM SURVEILLANCE VIDEOS: A CASE STUDY IN PREFABRICATED CONSTRUCTION

Abstract

Understanding the state, behavior and surrounding environment of construction workers is important for site management. Surveillance cameras are widely applied to monitor the working conditions of construction site. However, it is hard to automatically detect activities of every construction worker in real time from surveillance videos due to its far-field view and complex site situation. In order to address this problem, this thesis has proposed a valid vision-based technique for real-time activity recognition of construction workers. Deep learning replaces manual observations in this three-stage method for automatic monitoring using established on-site datasets. Surveillance videos with different degrees of recognition difficulties recorded from a prefabricated construction site are tested to prove the feasibility of this technique. This framework enables comprehensive visual understanding of construction site and contributes to productivity management with automatic monitoring of workers in real time.