Detection of story segments in news video footages

• Main Author: Octopios, Chowrichzy Wisdomy
• Other Authors: Teoh Eam Khwang
• Format: Final Year Project
• Language: English
• Published: 2011
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/42879
• Institution: Nanyang Technological University

Audio-visual content analysis is an area that is receiving increased interest, especially with the advent of large scale repositories of video content available through television, Internet and other channels of media distribution, for generating metadata that powers better search and recommendation. A core component of content analysis is the ability to segment the content into semantically disjoint parts/concepts. This project proposes a system for segmenting long news footages into its constituent semantic stories. A news story consists of several distinct semantic concepts, detecting which is easier for humans but not for a machine. Based on some apriori knowledge of how broadcast news is presented, detecting some important concepts can give us a clue to finding the story boundaries. Detecting the anchor person can give us some clue whether a news story has begun. While this may not be a new news story always, nevertheless this visual clue is quite crucial in finding a news story boundary. Finding the anchor person in news footages reliably is the primary goal on this project. A comparative analysis of the use of binary support vector classifier and one-class classifier is conducted to ascertain the suitability of a classifier for detection. Generating ground truth data for evaluating the performance of story boundary detection is an expensive and time consuming experience. In this project a user friendly annotation tool that is platform independent (runs on Win, Unix, and Mac) and allows detailed concept annotation is developed. This facilitates both easy collection of training and ground truth data for doing more extensive concept detection research. Based on the simulations done on four 30-minute Channel News Asia videos, the story boundary segmentation algorithm shows promising results. Further improvement on the system can be done by adding audio and text analysis.