Action recognition based on dynamic mode decomposition

Abstract Based on dynamic mode decomposition (DMD), a new empirical feature for quasi-few-shot setting (QFSS) skeleton-based action recognition (SAR) is proposed in this study. DMD linearizes the system and extracts modes form of flattened matrix or stacked eigenvalues, named feature. The has three advantages. first advantage its translational rotational invariance with respect to change localization pose camera. second one clear physical meaning, that is, if skeleton trajectory was treated as output nonlinear closed-loop system, then represent intrinsic property motion. Finally, last compact length simple calculation without training. information contained by not complete extracted using deep convolutional neural network (CNN). However, can be concatenated CNN features greatly improve their performance QFSS tasks, which we do have adequate samples train directly numerous support sets standard few-shot learning methods. Four datasets SAR CMU, Badminton, miniNTU-xsub, miniNTU-xview, are established based widely used public validate A group experiments conducted analyze properties DMD, whereas another focuses auxiliary functions. Experimental results show most typical tasks.