DL-Based Floorplan Generation from Noisy Point Clouds

Remote inspection of unknown and hostile environments can be performed by military/police personnel via the deployment of sensors and SLAM-based 3D reconstruction techniques. However, the generated point clouds cannot be transmitted to coordinators in real time, because of their large volume sizes. A common data-reduction solution is to convert 3D point cloud models into 2D floor plans. In this paper, we propose an end-to-end network for automated floor plan generation from noisy point clouds to estimate the main building structures (doors, windows and walls). First, the noisy 3D point cloud is column-filtered to remove irrelevant or noisy points. Second, we project the remaining points onto a grid map. Finally, an end-to-end neural network is trained to generate an accurate line-based floor plan from the grid map. Experimental results reveal that the proposed method generates floor plans that accurately represent the main structures of a building. On average, the estimated floor plans reach a 0.66 F₁ score for the building-layout evaluation, which outperforms the state-of-the-art methods. Furthermore, using floor plans reduces the model size by thousands of times on average, which enables real-time communication about the building structure.