An action can be represented as a sequence of salient postures. Effective modeling of the salient postures is critical for robust action recognition. This paper proposes to characterize the salient postures using a set of the spatiotemporal interesting points (STIPs). Local features are extracted at each STIP and the statistical distribution of the features for each salient posture is further modelled by a Gaussian mixture model (GMM). Experimental results have verified the effectiveness of the proposed posture model. © 2009 IEEE.