A RELIABLELOGO AND REPLAY DETECTOR FOR SPORTSVIDEO

Replay isoneofthekeycuesindicating highlights insports videos. A replay isusually sandwiched bytwoidentical logos whichprompt thestart andendofareplay. A logo transition usually contains 10-30frames, describes aflying orvarying object(s). Inthis paper, areliable logoandreplay detecting approach isproposed. Itcontains twomainstages: first, alogotransition template isunsupervised learned, a keyframe (K-frame) andasetofpixels that describes logo object (logo pixels, L-pixels) accurately arealsoextracted; second, thelearned information areusedjointly todetect logos andreplays inthevideo. Inaddition totraditional color analysis, optical flowfeature isemployed todepict the movementofthelogoobject(s). Extensive experiments showthat theproposed approach canreliably detect logos andreplays regardless ofthetypes ofsports videos.