Given the broad range of applications from video surveillance to humanâ??computer\ninteraction, human action learning and recognition analysis based on 3D skeleton data are currently\na popular area of research. In this paper, we propose a method for action recognition using depth\nsensors and representing the skeleton time series sequences as higher-order sparse structure tensors\nto exploit the dependencies among skeleton joints and to overcome the limitations of methods that\nuse joint coordinates as input signals. To this end, we estimate their decompositions based on\nrandomized subspace iteration that enables the computation of singular values and vectors of large\nsparse matrices with high accuracy. Specifically, we attempt to extract different feature representations\ncontaining spatio-temporal complementary information and extracting the mode-n singular values\nwith regards to the correlations of skeleton joints. Then, the extracted features are combined using\ndiscriminant correlation analysis, and a neural network is used to recognize the action patterns.\nThe experimental results presented use three widely used action datasets and confirm the great\npotential of the proposed action learning and recognition method.
Loading....