Recognizing Planned, Multiperson Action

摘要: Multiperson action recognition requires models of structured interaction between people and objects in the world. This paper demonstrates how highly structured, multiperson can be recognized from noisy perceptual data using visually grounded goal-based primitives low-order temporal relationships that are integrated a probabilistic framework.The representation, which is motivated by work model-based object plan recognition, makes four principal assumptions: (1) goals individual agents natural atomic representational units for specifying engaged group activities, (2) high-level description structure small set logical constraints adequate representing agent multiagent (3) Bayesian networks provide suitable mechanism integrating multiple sources uncertain visual feature evidence, (4) an automatically generated network used to combine information compute likelihood trajectory particular action.The method tested database American football play descriptions manually acquired but player trajectories. The strengths limitations system discussed compared with other algorithms.