视频内容的分析与理解往往基于对视频中目标对象的空间、运动特征进行感知。然而，在实际拍摄的视频中，目标对象的真实运动轨迹往往受到同时存在的相机全局运动影响。这种由相机自运动带来的全局运动在当前流行的自媒体视频中十分普遍。为了消除全局运动对视频中对象的真实运动轨迹的影响，提出了一种基于L1范数最小化的全局运动参数估计算法，并在此基础上实现了视频的全局运动补偿，得到了前景对象的真实运动轨迹。实验结果表明该算法能准确有效地去除全局运动的影响并准确恢复出运动对象的真实运动轨迹。

In the field of video content analysis and understanding, it is important to make good perception of spatial and motion features of objects in the video. However, in practice, object movement often mixes up with the camera movement which conceals the objects’ true trajectories. The camera movement, also named as globalmotion, which is induced by the movement of camera, is ubiquitous in the current popular we-media videos. To reveal the true object motion from the trajectories obtained from tracking algorithms, we propose a global motion estimation algorithm based on L1-norm minimization, and obtain the real motion trajectory of the foreground object, where the global motion compensation is realized. Experiments show that our algorithm could accurately estimate the inherent various global motions and restore the true motion trajectories effectively.

TP301.6

教育部人文社会科学研究青年项目（17YJCZH043）；重庆市教委科学技术研究项目（KJ1600937）。