A REAL-TIMEANALYSISOF 3DSCENEFROM MONOCULAR IMAGES BY OBSERVINGKNOWN BACKGROUND

Inthis paper, wepropose areal-time method foranalyzing 3-dimensional(3D) sceneofobjects fromtimeseries of monocular images. A single camera observes somepoints ontheobject sothat 3Dposition ofthepoints canbe estimated byextended Kalmanfilter. We applythis methodtotworeal-time applications. Oneistoacquire 3Dgeometry ofanobject, andanother istoestimate 3D pose/position ofanobject. Thisapproach needs nomodel dataoftheobject apriori andachieves theestimation of 3Dgeometry andpose/position. Intheexperiments using amanipulation robot, weshowthat itiseffective method toestimate 3Dinformation withhigh accuracy.