Abstrakt
The Euclidean metric, one of the classical similarity measures applied in clustering algorithms, has drawbacks when applied to spatial clustering. The resulting clusters are spherical and similarly sized, and the edges of objects are considerably smoothed.
This paper proposes a novel hybrid constrained pseudometric formed by the linear combination of the Euclidean metric and a pseudometric plus penalty. The pseudometric is used in a new deterministic incremental heuristic facility location algorithm (IHFL).
Our method generates larger, isotropic, and partially overlapping clusters of different sizes and spatial densities, better adapting to the surface complexity than the classical non-deterministic clustering. Cluster properties are used to derive new features for supervised/unsupervised learning.
Possible applications are the classification of point clouds, their simplification, detection, filtering, and extraction of different structural patterns or sampled objects. Experiments were run on point clouds derived from laser scanning and images.