Near Real-Time 3D Reconstruction of Construction Sites Based on Surveillance Cameras

The 3D reconstruction of construction sites is of great importance for construction progress, quality, and safety management. Currently, most of the existing 3D reconstruction methods are unable to conduct continuous and uninterrupted perception, and it is difficult to achieve registration with real coordinates and dimensions. This study proposes a hierarchical registration framework for 3D reconstruction of construction sites based on surveillance cameras. This method can quickly perform on-site 3D reconstruction and restoration by taking surveillance camera images as inputs. It combines 2D and 3D features and does not need transfer learning or camera calibration. By experimenting on one construction site, we found that this framework can complete the 3D point cloud estimation and registration of construction sites within an average of 3.105 s through surveillance images. The average RMSE of the point cloud within the site is 0.358 m, which is better than most point cloud registration methods. Through this method, 3D data within the scope of surveillance cameras can be quickly obtained, and the connection between 2D and 3D can be effectively established. Combined with visual information, it is beneficial to the digital twin management of construction sites.