Action recognition based on global optimal similarity measuring

Human action recognition based on the 3D skeleton is an important yet challenging task, because of the instability of skeleton joints and great variations in action length. In this paper we propose a novel method that can effectively deal with unstable joints and significant temporal misalignment. Action recognition is elegantly formulated as a sequence-matching problem on a pre-constructed weighted graph, which can encodes any spatio-temporal features and the transition probabilities between action elements. To classify any input sequence of actions, a global optimal matching algorithm based on dynamic programming is introduced, which can deal with temporal misalignment without pre-segmentation, The weighted graph is constructed in training stage. The proposed approach is evaluated on two benchmark datasets captured by a single depth sensor. Experimental results show that our approach can achieve superior performance to most state-of-the-art algorithms.