Human Tracking with Mixtures of Trees

Tree-structur ed probabilistic modelsadmit simple, fast inference. However, they are not well suitedto phenomena such asocclusion,where multiplecomponentsof an object maydisappearsimultaneously . Mixturesof treesappearto addressthis problem, at the cost of representinga large mixture. We demonstr ate an efficient and compactrepresentationof this mixture, which admitssimplelearningand inferencealgorithms. We usethis methodto build an automatedtracker for Muybridge sequencesof a variety of human activities. Tracking is difficult, becausethe temporal dependencies rule out simple inferencemethods. We showhow to use our modelfor efficient inference, usinga methodthat employsalternatespatial and temporal inference. Theresult is a tracker that (a) usesa veryloosemotionmodel,andso cantrack manydifferentactivitiesat a variable framerate and(b) is entirely automatic.