Floor Plan Generation via Ceiling Segmentation in Indoor Environmen

Abstract

In this paper, we present a novel approach for generating floor plans using odometry and partial observation segmentation from a mobile robot. With the advancements in robot manipulation and computer vision, mobile robots, such as robot vacuums, are increasingly employed in indoor environments. However, when using a restricted monocular camera as the sensor, constructing a map through visual simultaneous localization and mapping (SLAM) or visualizing the robot's path using odometry becomes challenging for users. To address this issue, we propose a floor plan generation method that leverages a ceiling segmentation model and the camera's intrinsic properties. In Chapter 1, we provide an overview of related work and outline the objectives of our study. In Chapter 2, we introduce a semantic segmentation model specifically designed for ceiling segmentation. This model utilizes sequential images to estimate the ceiling mask. Chapter 3 presents a postprocessing technique that involves projecting the output of the ceiling segmentation and the odometry onto the image. This process generates stacked masks, which are then refined using a refinement network. By incorporating the intrinsic matrix to unproject using homography and the wheel odometry, this method localizes the viewpoint within the indoor environment. By designing a hierarchical decoder estimates various components of floor plan including points, contours, and masks. Integrating the proposed methodologies from Chapter 2 to Chapter 3, we devise a comprehensive floor plan generation method that utilizes partial segmentation, postprocessing with image projection, and refinement. As a result, this research aims to facilitate the generation of user-friendly maps even in scenarios with limited sensors.