This tool can be used to segment a LiDAR point cloud based on differences in the orientation of fitted planar surfaces and point proximity. The algorithm begins by attempting to fit planar surfaces to all of the points within a user-specified radius (radius
) of each point in the LiDAR data set. The planar equation is stored for each point for which a suitable planar model can be fit. A region-growing algorithm is then used to assign nearby points with similar planar models. Similarity is based on a maximum allowable angular difference (in degrees) between the two neighbouring points' plane normal vectors (norm_diff
). The norm_diff
parameter can therefore be thought of as a way of specifying the magnitude of edges mapped by the region-growing algorithm. By setting this value appropriately, it is possible to segment each facet of a building's roof. Segment edges for planar points may also be determined by a maximum allowable height difference (maxzdiff
) between neighbouring points on the same plane. Points for which no suitable planar model can be fit are assigned to 'volume' (non-planar) segments (e.g. vegetation points) using a region-growing method that connects neighbouring points based solely on proximity (i.e. all volume points within radius
distance are considered to belong to the same segment).
The resulting point cloud will have both planar segments (largely ground surfaces and building roofs and walls) and volume segments (largely vegetation). Each segment is assigned a random red-green-blue (RGB) colour in the output LAS file. The largest segment in any airborne LiDAR dataset will usually belong to the ground surface. This largest segment will always be assigned a dark-green RGB of (25, 120, 0) by the tool.
This tool uses the random sample consensus (RANSAC) method to identify points within a LiDAR point cloud that belong to planar surfaces. RANSAC is a common method used in the field of computer vision to identify a subset of inlier points in a noisy data set containing abundant outlier points. Because LiDAR point clouds often contain vegetation points that do not form planar surfaces, this tool can be used to largely strip vegetation points from the point cloud, leaving behind the ground returns, buildings, and other points belonging to planar surfaces. If the classify
flag is used, non-planar points will not be removed but rather will be assigned a different class (1) than the planar points (0).
The algorithm selects a random sample, of a specified size (num_samples
) of the points from within the neighbourhood (radius
) surrounding each LiDAR point. The sample is then used to parameterize a planar best-fit model. The distance between each neighbouring point and the plane is then evaluated; inliers are those neighbouring points within a user-specified distance threshold (threshold
). Models with at least a minimum number of inlier points (model_size
) are then accepted. This process of selecting models is iterated a number of user-specified times (num_iter
).
One of the challenges with identifying planar surfaces in LiDAR point clouds is that these data are usually collected along scan lines. Therefore, each scan line can potentially yield a vertical planar surface, which is one reason that some vegetation points may be assigned to planes during the RANSAC plane-fitting method. To cope with this problem, the tool allows the user to specify a maximum planar slope (max_slope
) parameter. Planes that have slopes greater than this threshold are rejected by the algorithm. This has the side-effect of removing building walls however.
Fischler MA and Bolles RC. 1981. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381–395.
lidar_ransac_planes, lidar_ground_point_filter
def lidar_segmentation(self, in_lidar: Lidar, search_radius: float = 2.0, num_iterations: int = 50, num_samples: int = 10, inlier_threshold: float = 0.15, acceptable_model_size: int = 30, max_planar_slope: float = 75.0, norm_diff_threshold: float = 2.0, max_z_diff: float = 1.0, classes: bool = False, ground: bool = False) -> Lidar: ...