Event detection in soccer videos using shot focus identification

This paper presents an automatic event detection system fusing low and mid level features for soccer videos. We first employ an improved approach for Shot Boundary Detection with color and our mean-gradient feature. Then we classify the shots into two view types. We also perform a template-based replay detection for each shot. Play-break sequences are then generated using a rule-based method. We devise a novel method to locate the spatial position of event occurrence accurately using shot focus identification and extract other features from the play-break sequences. These descriptors are finally fed into a multi-kernel SVM classifier. Experiments tested on various matches demonstrate the effectiveness of our method.