An improved diffusion maps method for action recognition using global and local constraints

Action recognition is an important research issue in intelligent surveillance and many other automatic video systems. In this paper, we describe a novel method for the human action recognition from its silhouette in the video. In the algorithm, diffusion maps is used for dimensionality reduction as well as to preserve much of the geometrical structure. A global geometry and local temporal similarity is proposed to recognize the feature trajectory of actions in the learned eigen-space. The classification is performed in K-nearest neighbor framework. Extensive experiments on various scenarios from open databases are presented to demonstrate its high performance and strong robustness in comparison with previous algorithms.