A segment is characterized as being homogeneous. Homogeneity is mainly determined by the geometric constraints placed on point neighborhoods.
See also: Hoffman and Jain, 1987; Jain and Dubes, 1988; Filin 2002; Filin & Pfeifer 2006; Roggero 2002
Thus homogeneity can be defined by geometric point characteristics in accordance with neighborhood, radiometric homogeneity or larger topological recurring details. Each segment is a closed subset of the point cloud, V , and the segmentation operation will determine the character and number of segments. Every point in the cloud belongs to a segment, and no two segments have points in common.
Common techniques used for demarcating the feature space include K-Nearest Neighbour (KNN), and Maximum Likelihood methods (Jain and Dubes, 1988; Jain et al., 1999, 2000).
The characteristics of ALS data influence segmentation in two ways:
1. the resolution of data – In low resolution ALS data objects lose definition and therefore homogeneity constraints have to be relaxed. High resolution ALS data suffer from low amplitude and high frequency noise, which offers problems for homogeneity constraints based on measures like curvature.
2. Overlapping surfaces – In the real world objects sometimes overlap each other or the bare earth. For example, bridges overlap the bare earth and the branches of trees hang over roofs. Because of this neighborhood systems that are based on planimetric proximity will show great variety (gradient, curvature, etc.,) in the overlap regions. Therefore, neighborhood systems should ideally be based on proximity in 3D and hence allow for segments that overlap in planimetry.
(Silthole 2005, 67)