动态场景视频拼接方法研究

Video stitching for dynamic scenes is a challenging problem due to the existence of visible artifacts such as seams, blurring and ghosting. The core reason leading to these undesirable issues is multi-view parallax caused by insufficient accuracy of image alignment. This project proposes a coarse-to-fine registration approach including global registration, local alignment and pixel-wise adjustment. Introducing of inter-frame consistency enables the seamless stitching of videos shooting a dynamic scene. The study content contains: (1)The image alignment method based on localized facet fitting; (2) The structural misalignments eliminating based on variational local adjustment; and (3) video stitching under the constraints of inter-frame consistency. The characteristics and contributions are three folds: (1) Base on intrinsic and external parameters of cameras obtained from calibration, the robust and accurate image alignment can be achieved by localized facet fitting; (2) The disparity information is acquired by image matching and then propagated to the whole image plane using a variational method, the alignment precision could then be improved to pixel level; (3) The consistency of successive video frames is ensured by taking inter-frame constraints into consideration, and the computational speed could be accelerated at the same time. The achievement of this project would enrich and promote the researches on video stitching both practically and theoretically.

动态场景的视频拼接面临存在拼缝、模糊和鬼影等视觉失真的问题，其核心原因在于：多路摄像机之间一般存在视差，在场景动态变化的情况下，图像间的配准精度不够。本项目提出一种包含全局标定、局部拟合、像素级调整三个步骤的渐进式配准方法，引入帧间一致性约束，实现动态场景下视频的无缝拼接。研究内容：（1）基于局部平面拟合的图像配准方法；（2）基于局部调整的结构偏差消除方法；（3）帧间一致性约束下的视频拼接方法。特点和创新是：（1）充分利用摄像机固有的内参数和外参数，采用局部平面拟合方法，实现高鲁棒性、较高精度的图像配准；（2）基于图像匹配和变分理论，将偏差信息平滑扩散到整个图像范围，使得配准精度达到像素级；（3）引入帧间约束，保证视频中连续帧之间拼接结果的一致性，同时加速拼接算法。研究成果对丰富视频拼接理论、促进视频拼接技术的广泛应用具有重要意义。

动态场景的视频拼接面临存在拼缝、模糊和鬼影等视觉失真的问题，其核心原因在于：多路摄像机之间一般存在视差，在场景动态变化的情况下，图像间的配准精度不够。本项目提出一种包含全局标定、局部拟合、像素级调整三个步骤的渐进式配准方法。在三方面展开了研究工作：（1）构建了四镜头全景相机实验平台；（2）基于局部平面拟合的图像配准方法，充分利用摄像机固有的内参数和外参数，采用局部平面拟合方法，实现高鲁棒性、较高精度的图像配准；（3）基于局部调整的结构偏差消除方法，基于图像匹配和变分理论，将偏差信息平滑扩散到整个图像范围。研究成果对丰富视频拼接理论、促进视频拼接技术的广泛应用具有重要意义。