基于多域分析和全局优化的全景图拼接方法

为了快速鲁棒地进行全景图的拼接，提出了一种基于频率域和空间域特性的全局优化全景图拼接方法。该全景拼接过程分为局部对齐和全局对齐两个阶段。在局部对齐过程中，先使用相位相关法进行粗对齐，而在根据相位相关确定的重叠区域内部，则通过设计一种基于特征点的粒子滤波器来校正相位相关的结果；在全局对齐中，则是使用基于光差的全局优化方法在子像素级进行精确对齐，由于初值准确，因此全局算法能迅速收敛。但是由于全局优化会带来较大的参数空间，为此需使用小局部大全局的策略来降低参数空间。实验表明，该拼接方法在帧间光照变化较大、重叠区域较少、摄像机没有标定的情况下能够鲁棒地完成全景图的拼接。

Panoramic Mosaicing Based on Multi-domain Analysis and Global Optimization

In this paper a novel framework of panoramic mosaicing is presented based on phase correlation,particle filter and intensity difference minimizing. Combining the characteristics of frequency domain and spatial domain we construct a panorama from un-calibrated images for global optimization. The alignment consists of two phases: local alignment and global alignment. In the process of local alignment the lower accuracy,the proposed method employs phase correlation and feature based particle filter sequentially,by which we can obtain the swiftness and robustness of phase correlation as well as the corrective function of feature based particle filter. In the process of global alignment,because the initial value generated by the local alignment is close to the optimum one,the iterated algorithm could converge quickly. Meanwhile,a huge parameter space might be introduced by global optimization. We develop a strategy to reduce the dimensions of parameter space. In the experiment,this system shows the efficiency and robustness in the case of varying illumination,without camera calibration,less overlapping and less knowledge of the scene.