中文摘要：

为有效地保护版权，提高大规模视频集的拷贝检测速度，提出一种完全实现在GPU上的基于增量式聚类的拷贝检测方法．对数据库中新增加的视频，首先调用GPU上的硬件解码单元对视频流解码，以实时的速度提取高维SIFT特征点；然后对特征点进行增量K—means聚类，以动态地反映数据库的变化，并根据聚类结果更新视觉关键词词典；再将每帧表示成归一化的词频向量；最后使用基于帧级别词频向量的时空顺序匹配法来判定查询视频是否为数据库中视频的拷贝．实验结果表明，该方法比原有的CPU实现方法整体提速最高达63倍．

英文摘要：

For effectiveness of privacy protection and efficiency of copy detection on large video datasets, a fully GPU-based incremental copy detection scheme is proposed. When a newly added video arrives into the database, GPU on-chip decoder is called for video stream decoding. At the same time, high dimensional SIFT features are extracted on the frame in real-time, which is followed by an incremental K-means clustering method responding to the dynamic database used to update visual words codebook. Then, each frame is represented with a visual words frequency vector. Finally, a spatiotemporal sequence matching method based on visual words representation at frame level is used to determine whether the query is a copy. Experimental results show that our GPU implementation achieves up to a 63 times speedup over the CPU version.