Marker-Assisted Structure from Motion for 3D Environment Modeling and Object Pose Estimation

Accurately modeling as-built environments and tracking moving objects’ poses are critical for many Architecture, Engineering, Construction, and Facility Management (AECFM) automation applications. Equally important are the reliability, operating range and cost efficiency of such solutions for their broad deployment in unstructured, dynamic, and sometimes featureless AECFM sites. In this paper, a flexible vision-based technique is developed for accurate, robust, lowcost, and scalable pose estimation and as-built modeling in AECFM applications. This technique combines marker-based pose estimation and structure-from-motion (SfM). In the preparation phase, a sparse set of visual markers are installed in the target environment. During the operation phase, a set of unordered images are taken with a calibrated RGB camera. These images are immediately processed by a SfM system to estimate those markers' poses and generate a sparse point cloud, which can be used by robots or other mobile clients for either moving objects' pose estimation, or dimensional analysis of that environment. Furthermore, for as-built modeling, the RGB camera is replaced by a RGBD camera to create both a dense 3D point cloud and a concise planar model of the environment. Experiments have demonstrated sufficient accuracy (average absolute error within 5mm over a 9m scale) of the proposed technique. INTRODUCTION 3D geometric modeling in either construction sites or built environment has attracted increasing research interests in AECFM due to its importance for various construction and maintenance activities, such as as-built documentation, interior design and facility management. No matter what sensors are used for such modeling, a fundamental step is to find out different sensor poses (positions and orientations) in a same coordinate frame so as to reach a unified and meaningful result from raw data captured under different local coordinate frames of sensors. This is closely related to object pose estimation, another core problem appearing in many AECFM automation applications, such as safety and productivity monitoring of construction machinery. Among different technologies for 3D modeling or pose estimation, computer vision based methods have been introduced and investigated recently for potential construction applications. Whether the end result is a 3D model or an object's pose,