Event detection in sports video based on two-level semantic granularity extraction

In this paper, we present an approach for event detection in sports video based on semantic granularity extraction. This is accomplished by modeling two-level semantic granularity for video spatial and temporal content. Specially, in the stage of fine-grained segmentation, characteristics of salient regions serve as direct cues to identify temporal boundaries, and then spatio-temporal salient grain is extracted by temporal coherence analysis for grouping homogenous salient regions. Furthermore, in a coarse-grained manner, two classic semantic events Play/Break in sports video are detected using HMMs to combine prior knowledge with spatio-temporal descriptors of salient grains. The experimental results are presented to demonstrate the efficiency and accuracy of our proposed algorithm.