Learning and recognizing human dynamics in video sequences

摘要: This paper describes a probabilistic decomposition of human dynamics at multiple abstractions, and shows how to propagate hypotheses across space, time, abstraction levels. Recognition in this framework is the succession very general low level grouping mechanisms increased specific learned model based techniques higher Hard decision thresholds are delayed resolved by statistical models temporal context. Low-level primitives areas coherent motion found EM clustering, mid-level categories simple movements represented dynamical systems, high-level complex gestures Hidden Markov Models as successive phases ample movements. We show such representation can be from training data, apply It example gait recognition.