Fisher-ratio-separability boosted binary tree of posterior probability SVMs with application to action recognition

Based on fisher ratio class separability measure, we propose two types of posterior probability support vector machines (PPSVMs) using binary tree structure. The first one is a some-against-rest binary tree of PPSVM classifiers (SBT), for which some classes as a cluster are divided from the rest classes at each non-leaf node. To determine the two clusters, we use the Fisher ratio separability measure. Accordingly, the second proposed method termed one-against-rest binary tree of PPSVMs (OBT), we separate only one class with the largest separability measure from the rest classes at each non-leaf node. Then, the procedures of both SBT and OBT are provided. Finally, we consider the problem of human action recognition based on depth maps adopting both proposed approaches. Simulation results indicate both methods gain higher classifying accuracy than those of canonical multi-class SVMs and PPSVMs. Besides, the decision complexity of the proposed SBT and OBT are reduced because they use the posterior probability and the Fisher ratio separability measure.