Automated and real-time scan-to-BIM through deep learning-based object detection

Building information models (BIMs) of as-built environments are in growing demand in many engineering applications such as progress monitoring in construction, building renovation, project management, energy simulation, defect detection, and virtual site inspection. Scan-to-BIM is a technique that converts point clouds of buildings into as-built BIMs. Automated Scan-to-BIM for non-Manhattan structures and multi-room buildings remains an industry-wide challenge due to complex indoor environments and high requirements for generating volumetric and object-level representations. Conventional approaches are based on multiple procedures in extracting geometric and semantic features independently that cannot fully exploit object-level features. The feature extraction of existing approaches is mainly based on representations at point-, line-, or surface-levels. It is vitally important to develop object-level feature reasoning and an end-to-end trainable pipeline in order to achieve precise and robust Scan-to-BIM. This thesis aims to develop fully automated and real-time Scan-to-BIM solutions by leveraging deep learning techniques. Three new Scan-to-BIM approaches based on object detection have been proposed in this thesis. The first one is a twostage 3D object detection method using region-based convolutional neural networks (R-CNN). The feature fusion between sparse 3D and 2D bird's eye view (BEV) spaces is investigated to improve the generality and efficiency of modeling building primitives. In order to address the difficulties of training label generation caused by largely overlapped building objects, a dual-channel network is developed with one channel detecting walls and the other channel detecting remaining categories. The experimental results on the SUNCG dataset achieved recognition accuracy of 85.79% and localization precision of 79.03%, which have increased by 12.75% and 5.71% over the latest benchmark, respectively. It took an average of 4.75 s per scene with a mean footprint of 471.936 m2. The second approach is a corner-aware detector, Cor-Det, that simultaneously learns both object-level and corner-level features through corner-based supervision using deformable convolutions. The local features around the corners of the objects are incorporated in order to decompose the object location precisely. In experiments on the S3DIS dataset, Cor- Det outperforms state-of-the-art benchmarks with 80.5% recognition accuracy and 88.9% localization precision. The average time spent in modeling a single room is 0.53 s.