A new spatio temporal descriptor is proposed for action recognition. The action is modelled from a beam of trajec-tories obtained using semi dense point tracking on the video sequence. We detect the dominant points of these trajecto-ries as points of local extremum curvature and extract their corresponding feature vectors, to form a dictionary of atomic action elements. The high density of these informative and invariant elements allows effective statistical action descrip-tion. Then, human action recognition is performed using a bag of feature model with SVM classifier. Experimentations show promising results on several well-known datasets.