Automatic Floorplan Reconstruction from RGB-D Images. In Data Science & Engineering Master of Advanced Study (DSE MAS) Capstone Projects

This project aims to develop approaches to automatically reconstruct an accurate floor plan of a house with multiple rooms using a LiDAR-enabled smartphone. Previous work in this area, such as those described in research papers on the models FloorNet, Floor-SP, and 4D Spatio-Temporal ConvNets, have guided us in formulating our approach. We propose to accomplish this task via a combination of human annotation and deep learning segmentation models. We have developed tools to facilitate rapid human annotation of key features (such as windows, walls, and corners). The human annotations will be complemented by annotations generated by a deep learning segmentation model that can identify and locate doors. Finally, we have designed a cloud computing architecture that can store these annotations and build a digital floor plan of the 3D scene.