CALIBRATINGHEAD POSEESTIMATIONINVIDEOSFOR MEETING ROOM EVENT ANALYSIS

Inthis paper, westudy thecalibration ofheadposeestimation instereo camera setting formeeting roomvideo event analysis. Headposeinformation infers thedirection ofattention ofthesubjects invideo, therefore isvaluable forvideo event analysis/indexing, especially inmeeting roomscenario. We aredeveloping amulti-modal meeting roomdataanalyzing system forstudying meeting roominteraction dynamics, in whichheadposeestimation isoneofthekeycomponents. As eachsubject inthemeeting roomcanbeobserved byapair of stereo cameras, wedo2Dheadtracking forthesubject ineach camera, andthe3Dcoordinate oftheheadcanbeobtained by triangulation. The3Dheadposeisestimated inoneofthe cameracoordinate system, wedevelop aprocedure toaccurately convert theestimated 3Dposeinthecamera coordinate system tothat intheworldcoordinate system. Intheexperiment, visualization oftheestimated headposeandlocation inworld coordinate system verifies thesoundness ofourdesign. Theestimated headposeand3Dlocation ofthesubjects inthemeeting roomallows further analysis ofmeeting room interaction dynamics, suchasF-formation, Floor-control[l], etc. IndexTerms-Videosignal processing, Stereo vision, Machine Vision