Pixel based segmentation
Region based range image segmentation estimates curvature at each pixel and clustered pixels with homogenous surface. Thus locating extremities and surface discontinuities by classifying them by characteristics as: jump, crease, curvature: Jump edges can be detected using standard edge operators designed for intensity images such as the gradient, Sobel, Kirsch, Laplacian of Gaussian, and Canny edge operators (Suk & Bhandarkar 1992). Detection of crease and curvature edges, however, requires specialized operators. Once edge pixels have been detected and classified, they are linked together to form surface discontinuity contours.
Pattern based techniques – A problem with this algorithm is choosing the right features and normalizing them to avoid bias. Feature based clustering in ALS avoids geometric modelling on the presumption that geometric models describing the data are not readily discernible. The problem with this is that it is difficult to establish the relationship between the segmentation criterion and the properties of the desired segments.
Edge based segmentation – While essentially a sound concept, edge based algorithms suffer when the data is noisy (e.g., low vegetation), which can lead to open boundaries, multiple edges, and disconnected edges. Additionally, to obtain appropriate closing boundaries, the data also has to be processed at appropriate scales. Automating the determination of this scale can be problematic. However, there are other aspects of edge based algorithms that make them unsuitable. Edge based algorithms cannot cope with overlapping surfaces in ALS data. This means that the edges of bridges and buildings can be under segmented. (Silthole 2005, 71)
Segmentation by region growing
The segmentation method is based on region growing and uses the neighborhood definition of k nearest neighbors for each point:
•similarity of normal vectors (a)
•distance of candidate point to the adjusting plane (r)
•distance between current point and candidate point (d)
On approaches of manually designed decision tree and the statistical classifier, following template, and monument detection can be used within crop mark detection.