基于音频内容的篮球比赛精彩镜头检测

Highlight Detection in Basketball Video Based on Audio Content

提出了基于音频内容的篮球比赛精彩镜头检测系统。由音频关键词检测和精彩镜头检测2个子系统组成。第1个子系统采用二叉树结构的多级支持向量机(SVM)分类器及SEFC-FDR方法检测5个关键词。第2个子系统提出事件的二级转换模型。对3场总时长约319 min的NBA比赛进行测试,得分的平均准确率和回检率分别为64.89%和86.21%,犯规分别为64.60%和66.86%。

In this paper, a highlight detection system based on audio content is proposed for basketball games. It includes two subsystems, audio key words detection and highlight event detection. Five audio key words are detected in the first subsystem. A multi- stage Support Vector Machine （SVM） classifier with binary tree structure is combined with Simple Excellent Feature Combination method and Fisher＇s Discriminate Ratio （SEFC-FDR） to construct the first subsystem. In the second subsystem, a two-stage transition model of events is proposed. There are three NBA games about 319 minutes used to evaluate the proposed system. The testing results show that the average detection accuracy and recall rate for score event is 64.89% and 86.21%, respectively; and then for foul event is 64.60% and 66.86%, respectively.