Multi-Sample Image-Based Material Recognition and Formalized Sequencing Knowledge for Operation-Level Construction Progress Monitoring

Abstract

This paper presents a new method for operation-level monitoring of construction progress using image-based 3D point clouds and 4D Building Information Model (BIM). Previous research on comparing point clouds to 4D BIM has proven the practicality of performing progress monitoring by occupancy-based assessment: detecting if BIM elements are present in the scene. Nonetheless, without appearance information, operation-level monitoring – formwork vs. concrete surfaces for concrete placement– is still challenging. By leveraging the interconnectivity of site images and BIM-registered point clouds, this paper presents a new method for densely sampling and extracting 2D patches from all site images from which BIM elements are expected to be visible. Our method reasons about occlusions in the scene and classifies the material in each image patch. By formalizing the sequencing knowledge of construction operations for progress monitoring purposes and using histogram-based representation for possible types of construction materials, our method can accurately detect the current state-of-progress for BIM elements in presence of occlusions. We introduce a new image dataset for material recognition, and present promising results on operation-level progress monitoring on an actual concrete building construction site. Our method addresses the challenges of working with non-detailed BIM or high-level work breakdown structures.