3.2.3. Removal of Outliers

The overall accuracy rates of segmenting beams and columns are low. This may cause the subsequent automatic modeling of components to be inconsistent with the current situation. To solve this problem, this study refers to the method of Torr and Zisseee [36]. The use of building structure characteristics to filter out erroneous point clouds is proposed, and boundary errors are avoided. Three factors are considered: (1) point-to-plane distance; (2) plane normal vector; and (3) maximum angular distance, which uses the range interval and the directionality of the point cloud to remove the erroneous point cloud. The method first identifies the farthest point. Then, this point is used as the center to find the point cloud within a certain radius. This point is evaluated using the maximum and minimum values of the plane coordinates of the point cloud within the radius. Error points are filtered out.

In the category of beams and columns that are difficult to classify, a method for direction evaluation is added to filter out erroneous points and thus improve the accuracy of beam and column models. After this preliminary filtering of error points, the outline of the component becomes visible. However, if the coordinates of the maximum endpoint value of the point cloud are used directly as a component range, the appearance of the model may differ from the real situation.

Therefore, in extracting the boundaries of elements such as columns and beams in this study, the vertical axis (Z axis) is used as the normal vector for the column, and the horizontal axes (X and Y) are used as normal vectors for the beam element. Then, the erroneous points of the 3D point cloud with the maximum angle are removed. This method considers the allowable value of the angle between the point cloud and the normal vector. The point cloud within the allowable range is retained; otherwise, it is eliminated.
