Abstract: A template match-based method was presented by using the silhouettes of image sequence as representative descriptors of human posture to achieve human actions recognition．First,human silhouette extraction was obtained by background subtraction and shadow elimination．A new silhouette descriptor based on R transform was defined,which is employed to transform the silhouette into a 1D distance signal．Hierarchical clustering method was proposed to extract key postures of human action．Then,key postures were coded as a template called action string．At last,dynamic time warping was proposed to measure the similarity of the templates and the test sequences．This algorithm was evaluated using six human daily actions.The experimental results show that this method can achieve the correct recognition rate above 85%.

Key words: Computer vision,Silhouette,Key posture,Clustering,Action recognition