Local Structure Preserving Using Manifold Regularization and Pairwise Constraints for Action Recognition

With the rapid development of Internet technology and smart devices, tremendous amounts of multimedia data (e.g. text, image, video, audio, etc.) are produced and uploaded online every day. Semi-supervised learning has been proved to be one effect and effective solution to manage the massive emerging multimedia, which usually leverages the performance by exploiting the local geometry of a small number of labelled and a large number of unlabeled samples. The representative local structure preserving methods include manifold regularization and pairwise constraints. In this paper, we propose a local structure preserving method that effectively integrates manifold regularization and pairwise constraints. Particularly, we construct a new graph Laplacian by combining the traditional Laplacian and pairwise constraints. The new graph Laplacian can better preserve the local geometry and then further boost the performance. Finally, we build new local structure preserving classifiers including kernel least squares and support vector machines. We conduct extensive experiments on Chinese Academy of Sciences - Yunnan University - Multimodal Human Action Database (CAS-YNU-MHAD) for action recognition, respectively. The experimental results demonstrate that the proposed algorithm outperforms the baseline algorithms.