This paper presents a new approach to recognize human postures in video sequences comparing two methods. We first describe these two methods based on 2D appearances. The first one uses projections of moving pixels on the reference axis. The second method decomposes the human silhouette into blocks and learns 2D posture appearances through PCA. Then we use 3D model of posture to make the previous methods independent of the camera position. At the end we give some preliminary results and conclude on the effectiveness of this approach.