Classifying aggressive actions of 3D human models using correlation based affinity propagation algorithm

Abstract

Unlike traditional clustering techniques like K-means, K-nearest neighborhood where the number of initial clusters is known apriori, in affinity propagation (AP) based approach the numbers of clusters present in a given dataset are chosen automatically. The initial cluster centers in AP are termed as exemplars which are initialized randomly, and their numbers are automatically set and refined with the progress in the iteration. The AP has improved accuracy than K-means and has much lower computational time than the evolutionary based automatic clustering approaches and density based clustering like DBSCAN. The original AP algorithm was based on Euclidean distance. Here a Correlation based affinity propagation (CAP) algorithm is introduced considering Pearson Correlation as affinity measure. Simulation results reveal that the CAP provides effective clusters than that achieved by AP for datasets having a large number of attributes. The proposed algorithm is applied to categorize aggressive and regular actions of 3D human models. Extensive simulation studies on fifteen cases reveal the superior performance of CAP on finding the exact number of clusters as well as numbers of points in each cluster close to that of the true partition.