Enabling Component-Based Progress Monitoring on Construction Sites Through Image-Based Computer Vision

Vision-based construction monitoring methods have improved on-site transparency. However, many point cloud-based techniques are complex and often involve an image-dependent reconstruction step, making them prone to uncertainties. Additionally, few address productivity insights at the construction activity level. This paper presents a novel computer vision approach for automating construction progress monitoring, extracting information directly from image data enhanced through as-built details. A PIDNet Semantic segmentation model was trained to identify cast-in-place concrete walls, columns, and slabs during panel, rebar, and concrete phases. The detected components were processed using averaging techniques to monitor element-specific progress. The resulting data was integrated with as-built models through geometric projections, forming the basis for a digital twin construction. Our method was deployed on two-month construction data, providing detailed progress information and demonstrating its robustness. Compared to previous methods, this approach effectively merges existing as-built models with comprehensive as-performed image data.