Silhouette border’s extremities can be considered as describing features for human postures. Using of extremities for posture representation, and hence action recognition, is a simple and time efficient method that can also reduce the detection error rate. But unfortunately, popular extremity extraction methods e.g. star skeleton methods, encountered with enormous faults. This paper proposes a new method for extremity extraction with minimum error rate. First of all, curvature is introduced as a feature that can be used for extraction of candidate extremities. Since, the curvature is a boundary feature, it is insensitive to posture situations. Two filtering steps are then applied for minimizing wrong features. Since some of features may be blocked by occlusion a prediction step based on previous frames is employed. Pose representation is then carried out using a multi layer perceptron neural network. Promising achieved result verified the effectiveness of the proposed method. Experimental results demonstrate the privileged performance of the proposed algorithm to extract the extremities and so the human posture against the best reported works
