基于音/视频特征的足球视频体育事件交互式检索方法

An Approach for Football Video Event Retrieval Based on Audio/Video

本文提出了一种交互式足球体育视频事件检索方法。在该方法中,首先从音频和视频中提取四种类型的特征,计算出它们的均值和标准差,并把这八个数据编码成一个染色体,建立与视频文件的索引。然后,利用交互式遗传算法实现足球体育视频事件的检索。首先,系统从数据库中随机地选取N个视频文件供用户观看与选择;然后,系统根据用户所选视频提取相应的染色体,并对这些染色体进行重组操作得到目标染色体;其次,把目标染色体与数据库中的所有染色体进行比较,利用欧式距离计算出它们的相似度,从中选取N个最相似的染色体对应的视频为下一代视频;最后,不断迭代上面的过程,直到得到用户想要的视频。通过对包含有400个视频事件的数据库的实验,证明该方法能够有效地检索足球视频数据库中的视频文件,准确率达到89%。

This paper proposed an interactive genetic algorithm （IGA） for football video events retrieval with multimodal features. Eight audio-visual features,i, e. , average and standard deviation of shot duration, motion activity, sound energy, and speech rate, were extracted from each video;they were encoded as chromosomes and indexed into search table. First, the proposed algorithm randomly selected N videos from the initial population of videos database, and the user selected what he （she） wanted in mind. Next, the associated chromosomes of selected videos were regarded as target chromosomes after crossover, and chromosomes in the database videos were compared based on Euclidean distances to obtain the most similar videos as solutions of the next generation. By iterating this process, a new population of videos was retrieved. This approach of retrieval shows about 89% of precision on the average over 400 videos.