基于合并准则和测地膨胀重构的视频目标分割

Video Object Segmentation Based on Merging Criterion and Geodesic Dilation

提出具有鲁棒性的时空域视频目标分割方案:先用改进分水岭算法和合并准则作空域分割,产生初始I-帧分割,接着选择测地膨胀的掩模和标记图像来重构目标,以目标的每一轮廓像素构造边缘块,投影至下一帧中提取闭合目标轮廓。经多个视频序列实验证明,算法是行之有效的,而且适用于快速运动和部分遮挡的目标分割。

This paper suggests a robust spatial-temporal segmentation scheme. Spatial partition in I-frame is first implemented by a modified watershed transform and a region merging based on a new criterion. Every VO region is filled in by a geodesic reconstruction. Edge blocks centered at each pixel of VO contours will then be projected to the next frame to get closed VOs in other frames. Experiments on multiple sequences show that the algorithm is feasible and it is applicable to the segmentation of high speed and partially-kept-out object.