Soccer Video Event Detection Via Collaborative Textual, Aural And Visual Analysis
Loading...
Date
2011-10
Authors
Abdul Halin, Alfian
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Soccer event detection deals with identifying interesting segments in soccer video via audio/visual
content analysis. This task enables automatic high-level index creation, which circumvents
large-scale manual annotation and facilitates semantic-based retrieval. This thesis proposes
two frameworks for event detection through collaborative analysis of textual, aural and visual
features. The frameworks share a common initial component where both utilize an external
textual resource, which is the minute-by-minute (MBM) reports from sports broadcasters, to
accurately localize sections of video containing the desired events. The first framework identifies
an initial estimate of an eventful segment via audio energy analysis. Visual semantic
features are then observed to further refine the detected eventful segment. The second framework
implements a ranking procedure where semantic visual features are firstly analyzed to
generate a shortlist of candidates. This is followed by aural or visual analysis to rank the actual
eventful candidate top-most within the respective shortlist. Both frameworks rely on uncomplicated
audio/visual feature sets, which is the main advantage compared to previously proposed
works. Furthermore, manually labeled data are not needed since audio/visual considerations
are based on automatically classified semantic visual features and low-level aural calculations.
Evaluation made over a large video dataset shows promising results for goal, penalty, yellow
card, red card and substitution events detection.
Description
Keywords
Soccer video event detection , via collaborative textual