Unsupervised Clustering Based Real-time Shot Boundary Detection for Live Broadcasting
Ning Su, Jun Zhang, Yana Zhang, Guoting Zhang · 2019
The development of mobile Internet greatly facilitates information communication. The mobile live broadcasting platform has made video-sharing popular. In a long live broadcast, it is necessary to generate a video summary or some pieces of news for secondary information spreading. In order to produce a video summary in real-time, shot boundary detection (SBD) of video stream is badly needed. Traditionally, shot boundaries are determined by a threshold of feature distance in adjacent frames. It is difficult to decide a threshold applied in kinds of media content with good performance. Another, researchers have proposed Support Vector Machines (SVM) to classify shot boundaries, whose model relies much on the characteristics of training data. In terms of live broadcasting, real-time signal processing is required. This paper proposes an Unsupervised Clustering based Real-time SBD method (UCR-SBD), in which consideration is given to both detection efficiency and accuracy. Experiments show that the F value of SBD reaches 97.24% and the detection speed is 17 ms per frame on average. Finally, the proposed algorithm has been successfully applied in a real-time news reporting system for the live broadcast of CUC TV station.