TIME SENSITIVEAND NON-TIME SENSITIVEFEATURE EXTRACTIONS INARABICSIGN LANGUAGE RECOGNITION
T. Shanablehl · 2007
Thisworkintroduces twonovel approaches tofeature extractions ofvideo-based Arabicsignlanguage gestures namely: motion representation throughmotionestimation and motion representation through motionresiduals. Intheformer, motion estimation isusedtocompute themotion vectors ofavideo-based gesture. Thevertical andhorizontal components ofsuchvectors arerearranged intointensity imagesandtransformed intothe frequency domain. On theother hand,ifmotionisrepresented through motion residuals thensuchresiduals arethresholded and transformed intothefrequency domain. Themotion information is thentemporally accumulated through either telescopic motion vector composition orpolar accumulated differences. Thefeature vectors areextracted fromtheaccumulated motioninformation. Thesuperiority oftheproposed feature extraction techniques is illustrated through comparisons withexisting work.