基于抽的象隐马尔可夫模型的运动行为识别方法.pdf

22 3 Vol. 22 No. 3 2009 6 PR AI J n 2009 * 钱 堃 马旭东 戴先中 ( 210096) . . , . , Rao-b lackw ellised . , , . , , (EM ), , TP 24 M otion A ctivity R ecogn ition Based on Abstract H idden M arkov M od el Q IAN K n, MA X -Dong, DA I X ian-Zhong (K ey Laboratory of Measurem ent and Control of Comp lex Systems of Engineering, M inistry of Education, School of Automation, Southeast University, anj ing 210096) ABSTRACT R ecogn ition of h m an motion activity is essential in home-care robotic systems. In th is paper, a probab ilistic approach is proposed for h man motion activity recognition based on the abstract h idden M arkov model ( AHMM ). The AHMM is a w el-l s ited h ierarch ical model for representing goal-d irected motions at d ifferent levels of abstraction. In th is model, the decision m aking process of agent is eq ivalent to an abstractM arkov decision process (M DP). A model learn ing method is presented based on expectation-maxmi ization algorithm to learn the observation model and the transition model respectively. M oreover, approxmi ate inference of theAHMM is achieved by singR ao-blackw ellised particle filters and thereby it enables efficient comp tation in recognizing motion patterns. U sing trajectories derived from a vis al track ing system, several indoor motion patterns are recogn ized. Expermi ental res lts validate the good performance of the proposed approach. K ey W ord s H idden M arkov Mode,l Motion Estmi ation, Expectation-M axmi ization ( EM ) A lgorithm, A pproxmi ate Inference, Motion Activity Recognition * 863( No. 2006AA040202, 2007AA041703) : 2008- 03- 24; : 2008-