Traditional posture modeling for human action recognition is based on silhouette segmentation, which is subject to the noise from illumination variation and posture occlusions and shadow interruptions. In this paper, we extract spatial temporal movement features from human actions and adopt unsupervised clustering method for salient posture learning. First, spatial-temporal interest points (STIPs) were extracted according to the properties of human movement, and then, histogram of gradient was built to describe the distribution of STIPs in each frame for a single pose. In addition, the training samples were clustered by non-supervised classification method. Moreover, the salient postures were modeled with GMM according to Expectation Maximization (EM) estimation. The experiment results proved that our method can effectively and accurately recognize human’s action postures.
Bitte loggen Sie sich ein, um Zugang zu diesem Inhalt zu erhalten
