Human Action Recognition Based on Random Spectral Regression

For solving the uncertain parameter selection, the highly spatiotemporal complexity and the difficulty of effectively extracting feature in manifold learning algorithm processing higher-dimension of human action sequence, human action recognition algorithm based on random spectral regression (RSPR) is presented. The algorithm has three steps. Firstly, according to uniform distribution of human action data in the manifold and the classification labels of human action, the weight matrix is built. This method overcomes the neighborhood parameter selection of the manifold learning algorithm. Secondly, by spectral regression, the spatial manifold based on frame is approximated, and the manifold mapping of unlabeled sample is obtained. At last, the feature of the temporal series is extracted in the spatial manifold based on frame, and then in Gaussian process classification the feature of human action is classified. The experiment has three parts. When RSPR tests the recognition of human action by leave-one-out crossvalidation in Weizmann database, the recognition rate reach 93.2%; comparing RSPR with locality preserving projection (LPP) and neighborhood preserving embedding (NPE), through extracting the statistical feature of temporal sequences RSPR shows better performance; in the test of walk action influenced RSPR displays better adaptability.