A very low dimension frame-level motion descriptor is herein proposed with the capability to represent incomplete dynamics, thus allowing online action prediction. At each frame, a set of local trajectory kinematic cues are spatially pooled using a covariance matrix. The set of frame-level covariance matrices forms a Riemannian manifold that describes motion patterns. A set of statistic measures are computed over this manifold to characterize the sequence dynamics, either globally, or instantaneously from a motion history. Regarding the Riemannian metrics, two different versions are proposed: (1) by considering tangent projections with respect to updated recursive statistics, and (2) by mapping the covariance onto a linear matrix using as reference the identity matrix. The proposed approach was evaluated for two different tasks: (1) for action classification on complete video sequences and (2) for online action recognition, in which the activity is predicted at each frame. The method was evaluated using two public datasets: KTH and UT-interaction. For action classification, the method achieved an average accuracy of 92.27 and 81.67%, for KTH and UT-interaction, respectively. In partial recognition task, the proposed method achieved similar classification rate as for the whole sequence using only the 40 and 70% on KTH and UT sequences, respectively.