A Novel Story Unit Segmentation Algorithm Avoiding Voice Cutting

Automatic story unit segmentation is an essential technique for content based video retrieval and summarization. A good video story unit has natural boundary in visual and acoustic perception, respectively. For audio boundary, voice cutting problem affects acoustic perception of story unit. In this paper, a novel story unit segmentation algorithm is proposed to avoid the voice cutting problem. The algorithm combines replay detection, view pattern matching and non-speech detection to segment story units. Firstly, a replay detection method is implemented to find highlight events in soccer video. Secondly, based on positions of replay clips, a Fine State Machine (FSM) is used to obtain rough starting points of story units. Finally, audio boundary alignment is employed to locate natural audio boundaries for acoustic perception. The algorithm is tested on several broadcast soccer videos and the story units segmented by algorithms with and without audio alignment are compared in acoustic perception. The experimental results indicate the proposed algorithm can improve the acoustic perception of story unit by reducing the voice cutting problem.