Mapping and Automatic Post-Processing of Indoor Environments by Extending Visual SLAM

Adequately represented indoor mapping is of great importance. For instance, the efficiency and safety of first responders gets significantly improved. Stereo SLAM algorithms achieve remarkable results, nevertheless some of their map representations have drawbacks. Point clouds can rarely be interpreted intuitively by human operators due to the absence of color and their sparsity. Nor do those point clouds represent a suitable input for possible post-processing. This paper contributes a supplementary enhancement tailored to visual SLAM algorithms. First, a method of creating dense colored point clouds is introduced, which is based on 3D reconstruction and visual SLAM algorithms. Second, a post-processing scheme is proposed, which condenses those point clouds to a blueprint-like map of the building by extracting wall segments. Evaluation is conducted on data recorded with a person-carried stereo camera setup in a typical indoor environment covering a track length of about 70 m. It is demonstrated that a dense colored point cloud has been successfully created and adds great value to intuitive interpretation by human operators. Furthermore, post-processing delivers promising results. A reliable extraction of wall segments is achieved for walls that are represented by a sufficient number of points.