Comparative Analysis of Action Recognition Methodologies using Construction Data in CCTV for Construction Sites

초록

Recent studies on vision-based equipment recognition and activity analysis have shown promising results on automated productivity assessment of construction operations. However, the analysis conducted in the existing studies rely on high-quality video data sets that is often not available in actual construction sites such as low quality of on-site CCTV footage. In this study, three representative deep learning models were compared for this matter using low-quality CCTV data of excavators collected from an on-going road construction site. The goal of this study is not to determine the best-performing model, but to identify the one that remains stable under low-quality video conditions. All clips were standardized to 150 frames to ensure consistent input for training. Experimental results show that BiLRCN achieved the highest accuracy (0.993) and stable learning performance. LRCN exhibited minor fluctuations during validation, while 3D-ResNet effectively captured spatiotemporal features but struggled with rotation-related actions. Overall, the bidirectional recurrent structure of BiLRCN demonstrated the most reliable performance under low-quality video environments, suggesting its potential applicability for automated productivity monitoring in real construction projects.

키워드