Abstract: A hidden conditional random fields-based method for action recognition was presented by using the contours of image sequence as representative descriptors of human posture. First, human silhouette contour extraction is obtained by background subtraction and shadow elimination. Star skeleton which uses the centroid-border distance function de- scribing the contour is not good at describing the local features. A new contour descriptor based on distance array was defined, which is employed to transform the contour into a 1D distance signal. At last, the discriminative hidden condi- tional random fields arc used to recognize human actions. I}he experimental results show that this method can achieve the correct recognition rate above 91. 4 0 0.

Key words: Contour,Discriminative model, Hidden conditional random fields,Action recognition