Box-Loc: Box-aware Module Localization Using Point Cloud and Deep Learning in Modular Integrated Construction

Three-dimensional (3D) volumetric modules are critical elements to be monitored and managed at modular integrated construction (MiC) sites. Intelligent module localization with excellent performance should be beneficial but has been underexplored. This study developed a box-aware module localization (Box-Loc) model by integrating point cloud and deep learning to get module locations from 3D bounding boxes. Synthetic datasets were established combining real-life and virtual prototyping-enabled pseudo-lidar point clouds. The model adapted 3D object detection algorithms to the construction context and employed transfer learning to mitigate overfitting. The real-life case study demonstrates that the Box-Loc model outperformed previous methods, reducing the average error from tens of centimeters to a minimum of 8.46 cm in the Y coordinate and decreasing the inference time from minutes to 55.7 ms. The real-time and centimeter-level module localization should facilitate safer and more efficient site monitoring and management, thus motivating a wider adoption of MiC globally.