111 thi.) puper, we present un uutomutic. system fi,r UIIulyzing and unnotuting video sequences of trchnicrd tulks. Our method uses U robust motion estiniution tec,hnique tci detect k e ~frunzes und segment the video sequence into subsequences cmtuining a single overhead slide. The subsequences ure stabilized to remove motion that o m mwhen the speukeradjusts their slides. Any chunges remuining betweenfrunzes in the stubilized sequencxs muy be due to speuker gestures such uspointing or writing and we use uc-tivec~intour.~ to automaticdl~truck these potential gestures. Given the construined domuin we dejine U simple "voc-ubu1un." cguctions which ~ ' u neusily be recognized bused on the active c~intourshupe and motion. The recognized actionsprovide U rich unnotution ofthe sequence that cun be used to ucces.sa condensed version of the tulkfrom U webpuge.
Shanon X. Ju, Michael J. Black, Scott L. Minneman,