基于内容的相似视频片段检索技术

Content-based Near Duplicate Video Clip Retrieval Technology

为了实现相似视频片段的快速检索,提出一种建立在视频单元层上的视频片段探测方法.在视频片段的特征描述阶段,利用视频距离轨迹(VDT)来表现视频片段随时间变化的趋势,同时定义线性拟合特征(LFF)的描述方法来描述VDT的特征.在特征匹配阶段,利用图论中的最优匹配方法寻找视频单元序列之间的最佳匹配方式,同时定义视觉相似度和时序相似度的计算方法,在保证匹配准确度的同时大幅度减少了匹配的数据量.实验结果表明,该方法极大地改善了相似视频片段的检索效果,与传统的视频检索方法相比,该方法具有更好的鲁棒性以及更加快速的效率.

In order to realize the fast retrieval of near duplicate video clip, a segment based video clip retrieval algorithm is presented in this paper. During the feature describing stage, the video distance track （ VDT } which is described by the linear fitting feature （ LFF } is used to mine the feature of video clip. During the feature matching stage, the optimized marching approach of Graph Theo- ry is employed to find the best alignment between video clip sequences. Meanwhile, the visual similarity and time sequence similarity＇ s defining are given. Experimental results indicate that these methods greatly improve the retrieval effect of video clips. Compared with traditional video retrieval methods, the proposed method are more robust and fast.