*2.1. Case Studies*

For the current research, two case study examples are used. The first case study is the building of the Technical University in Braunschweig (TUB) and the second is the University of Twente Faculty of Geo-Information Science and Earth Observation (ITC) building. The floor plans of these buildings are shown in Figure 4. In Figure 4a, the highlighted area shows that a wall was removed and rooms were merged into one, and Figure 4b shows the two rooms before removing the walls.

**Figure 4.** The floor plans for our two case studies. (**a**) (**top**) TU Braunschweig after the change (floor2), (**b**) (**bottom**) University of Twente ITC building (floor 1) before the change. The highlighted areas show the rooms where the changes happened.

Point cloud data for the two case studies were collected with different scanners (Figure 5). The data for the Braunschweig building were collected with an ITC Indoor Mobile Mapping System (ITC-IMMS) (epoch1) [43] and a Zeb-Revo (epoch2) [44]. For the ITC building, we used the Riegl [45] terrestrial laser scanning system and a Viametris device. The accuracy of the point clouds varied from 0.01 to 0.06 m depending on the laser scanner system. While the noise in mobile mapping systems was louder than the terrestrial laser scanner (TLS), the scene coverage of a mobile mapping system was more than a TLS. The noise in the data could have been caused by sensors, data acquisition algorithms, and the reflective surfaces. For more information on the comparison of scanning systems, refer to the study by Lehtola et al. [46].

**Figure 5.** The datasets for two different epochs. The first row is the dataset which belongs to the Braunschweig building and the second row is from the ITC building and is a more complex dataset, with furniture and large glass windows.

In the following subsections, the detailed methodology is explained based on the first case study.

#### *2.2. Indoor Change Detection from Point Clouds*

Differences in two epochs of point clouds inside the buildings can be categorized as:


There are some other differences between two epochs of point clouds that are interpreted as:


In our approach, categories number 1 and 2 were dealt with as important changes for 3D Cadastre, and categories 3 and 4 were just inevitable differences in two epochs that occurred because of data acquisition systems and were not relevant to the 3D Cadastre. We acquired two point clouds of two time periods with two different laser scanners, one a Zeb-Revo [44] handheld MLS and the other an ITC-IMMS [43]. The motivation to use different sensors is to explore all realistic possible causes of differences between epochs. The process of change detection starts with the co-registration of two point clouds (Step 3 from Figure 3). The co-registration of two point cloud datasets was a straightforward approach, such as using the iterative closest point (ICP) [47–49]. After the registration, two point clouds were compared based on the distance threshold to detect the differences caused by the registration error and sensors differences (4th category; Step 4 from Figure 3). The distance threshold was chosen by summing the registration error and sensor noise. The registration error and sensor noise already introduced some differences between the two datasets. The registration errors were the residuals of each co-registration process (less than 10 cm). The sensor noise was specified in the specification of the systems. This threshold, *d*, described points from two datasets with the distance less than the threshold. They were not considered as changes and they were in the 4th category because of the differences in the sensors. Points that had distances more than the threshold were in one of the other three categories. In our experiment, we defined the distance threshold of less than 0.10 m.

Let the point clouds (PC) from epoch one (acquired by a backpack) be PC1 and the point clouds from the second epoch (acquired by Zeb-Revo) be PC2. The point to point comparison was based on the reconstruction of a Kd-tree [50,51] and a comparison of the distance of the points in PC1 from PC2 and was stored in PC1. Using this method, the differences caused by the acquisition system and registration errors were excluded from the real changes.

In the next Step 5 from Figure 3, the differences were further analyzed to detect and exclude the acquisition coverage (3rd category). Our change detection method was based on analyzing two geometric differences between two point clouds. This was done in two steps: (1) The distinction was made between object changes and coverage differences and (2) the object changes were separated into changes on permanent structures and dynamic objects, such as persons and furniture (Section 2.2).

The geometric differences were calculated by determining the nearest 2D point and the nearest 3D point in the other epoch. The first nearest point was based on the X, Y coordinates and the second on X, Y, Z coordinates. Figure 6 shows both geometric distances as a point attribute categorized in three colors: Green <20 cm, yellow >20 cm and <50 cm, red >50 cm to the nearest.

**Figure 6.** The distance (green <20 cm, yellow <50 cm, red >50 cm) to the nearest point in (**a**) 2D and (**b**) 3D.

For both object changes and coverage differences, it was expected that the nearest 3D point was further than a certain threshold. However, the nearest 2D point may have been close to a changed object, but not in case of coverage differences. Points were temporarily labeled as part of changed objects if the distance to the nearest point in 3D was larger than 20 cm, but the nearest point in 2D was less than 20 cm. Threshold values were chosen such that they were larger than the expected registration errors but small enough to detect changes larger than 20 cm. Next, the whole point cloud was segmented into planar segments and only the vertical segments with a majority (more than 50%) of points labeled as potentially changed were considered to be changed. The planar segmentation was performed by a region growing algorithm presented by Vosselman et al., [52]. Note that, in this way, the points on a newly built wall near the ground or ceiling, with a small 3D distance to the nearest point in the other epoch, were included in the changed objects as they belonged to a segment with more than 50% points with a large perpendicular distance to the plane in the other epoch. By using planar segments and calculating perpendicular distances from a point in one epoch to a plane in the other epoch, we avoided the influence of differences in point densities between the point clouds. The vertical segments labeled as changed objects included permanent structures, such as walls, but also dynamic objects, such as persons. In the second step, the aim was to separate permanent from temporary changes by looking at a method described in [23] and [42].
