中文摘要：

无人机航拍视频中经常出现航向频繁改变的情况。基于条带的视频拼接算法尽管在处理速度上有很大优势,但是限制了相机的拍摄方式,主要处理平动相机拍摄的视频。用于采集全景图信息的纵向条带如用来扫描非平动场景会使效率降低甚至失效,所以不能直接拼接航拍视频。针对该问题,本文提出一种多向条带拼接算法,通过量化场景运动方向到预先定义的多个区间,分情况选取最佳方向的条带,控制了条带扫描方向与场景运动方向之间的最大夹角,从而避免了采集效率的降低。实验结果表明这种方法能有效拼接航向变化的仿真视频,全景图结果覆盖的场景范围大于原算法,证明了算法的有效性。

英文摘要：

The regular manifold projection method for video stitching is very efficient but it requires an invariant dominant motion.However,aerial videos captured on unmanned aerial vehicles cannot be stitched by it due to the flexibility in their flight routes.This paper proposes an improved method to handle the varying direction problems via multi-direction strips.First,various slope strips for different scan directions are defined.When the major motion direction in the aerial video changes,the most appropriate strip are chosen to minimize the bias angle between the scan direction and the motion direction.Experiments show our method can generate better panorama result as compared with other methods.