In this paper, we propose a new method for body gesture recognition based upon Laban Movement Analysis (LMA). The features are computed for a dataset of pre-segmented sequences putting at stake 11 different actions, and are used to build a dictionary of key poses, obtained with the help of a k-means clustering approach. A soft assignment method based upon the obtained poses is applied to the dataset and assignment results are used as input sequences in a Hidden Markov Models (HMM) framework for real-time action recognition purpose. The high recognition rates obtained (more than 92% for certain gestures), demonstrate the pertinence of the proposed method