A compact and recursive Riemannian motion descriptor for untrimmed activity recognition

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 covariance matrix. The matrices forms Riemannian manifold that describes patterns. statistic measures computed over this characterize sequence either globally, or instantaneously from history. Regarding metrics, two different versions proposed: (1) by considering tangent projections respect updated recursive statistics, and (2) mapping onto linear matrix as reference identity approach was evaluated for tasks: classification on complete video sequences recognition, in which activity predicted at frame. method public datasets: KTH UT-interaction. For classification, achieved an average accuracy 92.27 81.67%, UT-interaction, respectively. In partial recognition task, similar rate whole only 40 70% UT sequences, code work available [code].