K-center algorithm for hierarchical binary template matching

Abstract

To construct a hierarchical template tree of binary templates, the dissimilarity between templates was defined using distance transform, and a k-medoid algorithm was applied to select the representative of a template cluster. However, this method has a limitation in that the representative of a higher level cluster cannot exist in the space between the lower level clusters. In order to solve this problem, this paper proposes a k-center algorithm that finds a template that minimizes the distance to templates belonging to a cluster by using a genetic algorithm regarding each pixel of a binary template as a gene. The search space can be limited by randomly selecting the initial population from the templates belonging to the cluster. In particular, the weighted bidirectional distance is used as the dissimilarity between templates to prevent the cluster center from shrinking. By applying the proposed method to pedestrian silhouettes, it is possible to confirm that a leaf node similar to the given template is better detected and its performance is more stable.