Activity Recognition via Autoregressive Prediction of Velocity Distribution

We present a novel approach for view-based learning and recognition of motion patterns of articulated objects. We formulate the intervals of motion as a predictive model of local spatio-temporal receptive field activation. We compute local velocity distribution using a Bayesian approach, and then approximate the local velocity distribution in space and time using a set of Gaussian receptive fields. The activation sequence of receptive fields over time is modeled in a PCA subspace using linear auto-regression to arrive at a model of the motion pattern. Recognition is performed using the MDL principle. We test the approach on a number of human motion patterns to demonstrate the applicability of the proposed approach to simple action recognition and identification.