H．264压缩域中利用Biased-SVM检测镜头边界

为了直接从H．264码流中检测镜头边界，提出了利用H．264压缩域多特征和Biased—SVM（不平衡支持向量机）分类算法的检测方法。分析帧类型、宏块类型、运动矢量、帧内预测模式等信息，以获得发生镜头突变和渐变的特征。针对镜头边界帧的数量远少于视频帧总数的特点，用Biased—SVM分类方法将视频帧分为突变帧、渐变帧和非镜头边界帧。在TRECVID视频集上的实验结果表明，与其他H．264压缩域的算法相比，该算法有更好的性能。

Shot boundary detection using Biased-SVM in H.264 compressed domain

In order to detect shot boundaries in H.264 bit streams, a shot boundary detection method using compressed domain features of H.264 and BiasedSVM （Biased Support Vector Machine） is proposed. The features about the abrupt shot changes and gradual shot changes are obtained by analyzing the information of frame type, macroblock type, motion vector, intraprediction mode, etc. As the number of shot boundary frames is far fewer than the total number of video frames, proposed method chooses BiasedSVM to classify the frames into three classes, namely, the frames of abrupt change, gradual change and nonchange. Experi mental results on TRECVID video dataset indicate that the presented approach has better performance on shot boundary detection, compared with other method in H.264 compressed domain.