AUTOMATING INDOOR ELEMENT IDENTIFICATION AND GLOBAL POSITIONING: AN IMAGE RETRIEVAL-BASED BIM DEVELOPMENT SYSTEM FROM THE INTERSECTION OF VIRTUAL CAMERA VIEWS

Abstract

Scene recognition and localization services are increasingly necessary for better management of indoor facilities of buildings. Recent deep learning-based studies require much time for manual labeling and cannot also meet the demands of specific element information and global positioning. In this paper, an image retrieval-based BIM development framework is proposed to automatedly and simultaneously identify indoor elements and obtain global positioning. Specifically, OpenStreetMap (OSM) and BIM are combined to build a digital-based virtual environment. Based on its 2D images, style transfer and image retrieval are conducted by Convolutional Neural Networks (CNN). BIM development algorithms are then proposed to achieve the purpose by the overlap of virtual camera views from BIM and coordinate transformation based on physical world positioning data in OSM. The experiment demonstrated the effectiveness of the proposed method with good performance regarding identification accuracy, positioning error, and orientation error. It is also robust under different indoor environmental conditions.