Abstract

This paper focuses on human posture clustering and modeling for human action recognition in the field of computer vision. Specifically we mainly talk about posture description with spatial temporal interesting point features rather than traditional posture segmentations; also we give the comparisons of four kinds of unsupervised clustering methods and continue to carry out unsupervised posture classifications based on Weizmann database. In the following we use GMMs based on EM algorithm to model each clustered posture type. Finally we test our method with Weizmann and KTH Action database. These experiments show its effectiveness and robustness.