Resume These Mathis Baubriaud

Reporting automatisé de l’état d’avancement d’un chantier à l’aide de la réalité mixte

Author : Mathis Baubriaud

Abstract

Automated inspection of construction sites represents a significant opportunity to improve efficiency and accuracy in the construction industry. This thesis addresses the challenges of interior construction progress monitoring by developing and evaluating an innovative augmented reality (AR) solution. By overlaying Building Information Modeling (BIM) data onto real-world environments through an AR headset, the proposed system aims to automatically detect and identify HVAC and electrical components. These elements, with their heterogeneous shapes, present unique challenges for computer vision. For a complete analysis, two complementary methods are presented: one, geometric, relies on a ray-tracing algorithm and 3D reconstruction of point clouds acquired by depth sensors; the other, based on 2D image analysis, uses a deep neural network. A critical obstacle in applying deep learning to this domain is the lack of large, annotated datasets for training. To address this, the study presents an automated process for generating photorealistic synthetic datasets from BIM models using a graphics engine. These datasets enable the training of a state-of-the-art instance segmentation network, which is enhanced by transfer learning methods. Experiments conducted on real construction sites demonstrate the system’s potential for more efficient and reliable progress monitoring. The integration of synthetic data generation, deep learning, and AR demonstrates the added value of this approach, advancing current practices in indoor construction progress monitoring and highlighting its feasibility for real- world application.

Keywords : Building Information Modeling (BIM); Computer Vision; Augmented Reality; Deep Learning; Progress Monitoring; Indoor Construction.