Activity recognition via autoregressive prediction of velocity distribution.pdfVIP

Activity Recognition via Autoregressive Prediction of Velocity Distribution Miha Peternel and Ales? Leonardis University of Ljubljana Faculty of Computer and Information Science Trz?as?ka 25, SI-1001 Ljubljana, Slovenia {miha.peternel,ales.leonardis}@fri.uni-lj.si Abstract 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. 1 Introduction The ability to learn and later recognize articulated activities with few assumptions about the object geometry, appear- ance, and the nature of the activity would have a number of applications in monitoring, video interpretation, video indexing, smart environments, and human-robot interactions. Several models have been developed to enable recognition of motion patterns, trajectories, and consequently activ- ities [7]. Standard methods either use a predefined geometrical model and try to estimate its parameters [8, 1, 2, 11], or attempt to model the observed motion directly [3, 16, 4, 5]. Recently, attempts have been made to summarize motion using local spatio-temporal features [14, 12]. It has been shown [12] that distribution density of local trajectories can be used for learning and recognition of cyclic human motion, however such representation is scale dependent and ap- plies to cyclic motion only. Fablet et al. [6] have