Automatic reality capture and monitoring of construction sites can reduce costs, accelerate timelines and improve quality in construction projects. Recently, automatic close-range capture of the state of large construction sites has become possible through crane and drone-mounted cameras, which results in sizeable, noisy, multi-building as-built point clouds. To infer construction progress from these point clouds, they must be aligned with the as-designed BIM model. Unlike the problem of aligning single buildings, the multi-building scenario is not well-studied. In this work, we address some unique issues that arise in the alignment of multi-building point clouds. Firstly, we show that a BIM-based 3D filter is a versatile tool that can be used at multiple stages of the alignment process. We use the building-pass filter to remove non-building noise and thus extract the buildings, delineate the boundaries of the building after the base is identified and as a post-processing step after the alignment is achieved. Secondly, in light of the sparseness of some buildings due to partial capture, we propose to use the best-captured building as a pivot to align the entire point cloud. We propose a fully automated three-step alignment process that leverages the simple geometry of the pivot building and aligns partial xy-projections, identifies the base using z-histograms and aligns the bounding boxes of partial yz-projections. Experimental results with crane camera point clouds of a large construction site show that our proposed techniques are fast and accurate, allowing us to estimate the current floor under construction from the aligned clouds and enabling potential slab state analysis. This work contributes a fully automated method of reality capture and monitoring of multi-building construction sites.

Monitoring the progress of a large construction site manually is a challenging task for managers. By collecting visual data of the site, many monitoring tasks can be automated using machine vision techniques. In this work, we study a new method of collecting site data, which is through crane camera images used to create 3D point clouds. The technology is cost-effective and enables automatic capturing and transmission of on-site data. To automatically extract buildings from the as-built point clouds, we present VBUILT, which uses 3D convex hull volumes to identify building clusters. Experimental results on 40 point clouds collected over four months on a large construction site show that the proposed algorithm can identify building clusters with 100% accuracy.