TV commercial segmentation using audiovisual features and support vector machine

Commercial management is an emerging technology. Most of the previous works are focus on detection of commercial blocks, while ignoring the segmentation of commercials which is an primary step of funding the known commercial database. In this paper, we transform the problem to commercial boundary detection, and propose a novel method to fuse audio and visual features to find the individual commercial boundary. Audiovisual features such as FMPI (Image Frames Marked with Product Information) and speaker change probability are extracted for each shot boundary. And then, contextual features which are generated from these basic audiovisual features are used to predict the probabilities of commercial boundary for every shot boundary. At last, a post-processing method is utilized to refine the result. Experiments show a promising result on a commercial database from TV channels in China.