Development of Deep Learning Methodology of Object and Action Recognition for Discrete Event Simulation in Construction

초록

Enhancing the productivity and detecting inefficiency of earthmoving equipment are key factors leading to successful project delivery. This requires sufficient time and effort of data collection and simulation to optimize equipment fleet. This study proposes a deep learning–based framework that combines video-based action recognition with Web CYCLONE simulation for automatic activity time collection that is used as inputs of discrete event simulation model of earthmoving operation. CCTV footage collected from an on-going road construction site was collected. Excavators were detected, tracked, and classified into four actions—excavating, rotating, loading, and returning—through an integrated Faster R-CNN, SORT, and BiLRCN approach. The extracted action durations were applied to Web CYCLONE to estimate cycle time and productivity. Simulation results showed an average cycle of about 175 seconds, with the hauling phase occupying 85% of the total duration, confirming it as the dominant productivity factor. The proposed method demonstrates that automated video analysis can identify process bottlenecks and support data-driven productivity assessment in smart construction.

키워드