基于时域统计变化检测实时分割头肩视频序列

Real-time segmentation for head-shoulder video sequences based on temporal statistical change detection

分析了头肩视频序列的特点,提出了基于时域统计变化检测、利用多帧运动信息实时分割视频对象的方法。先选取包括当前帧在内的前连续2N帧图像,将奇数帧与偶数帧图像作差值,形成长度为N的帧差图像序列;对每个象素点时域上的N个帧差样本值进行t分布显著性检验,判断象素点是否发生了变化;对得到的二值图像进行形态学处理,得到完整的分割结果。试验结果表明,该算法能够自动实时的分割视频对象。

The characteristics of the head-shoulder video sequences were analyzed,and a new method for real-time segmenting video object using motion information from multiple frames was bought out based on temporal statistical change detection. First,2N successive frames ranging from the current frame n to frame n-2N+1 were selected,and N frame differences between the odd frame and the even frame were computed to form a frame difference sequence. Then,a t-distribution significance test was performed on the N temporal difference samples for every pixel to judge whether it was changed. Finally,morphology processing was performed on the binary image obtained from the previous step to get the segmentation result. Experiment results show that the new algorithm can real-time segment video objects automatically.