An automated semantic segmentation-free approach to point cloud scene understanding in construction

Abstract

It has been long shown that deep learning methods need voluminous training data to automatically process point clouds and develop accurate as-built construction models. On the other hand, conventional semantic segmentation of point clouds can be used but requires significant time and manual effort. In construction, public and private Building Information Model (BIM) databases are available as information source and can facilitate semantic registration and enrichment. Therefore, this paper takes the initial steps and advances the Derivative-Free Optimization-based (DFO) approach by automatically recognizing 3D objects from point clouds using available BIM data and registering the correlated semantic information to each detected object using geometry and color features. More specifically, the proposed framework takes the as-designed BIM models and point clouds as input, samples as-designed models, and registers information as a log text file to the generated point cloud by minimizing the Mean Squared Error (MSE) value using geometry and color features. In order to validate the proposed framework, preliminary experiments were conducted on an office point cloud taken from the Stanford 2D-3D dataset. Results highlighted the potential of the proposed DFO approach in making good use of available BIM resources to efficiently process point clouds and generate accurate and semantically enriched as-built models.