Abstract:  The current mainstream panorama stitching softwares heavily rely on the precision of matched points in street view images in the process of image warping. In order to enhance robustness, the strategy of getting feature points in warped images and transforming them to street view images is adopted. Besides, stitching seams between adjacent images are removed by nonrigid image warping, which is achieved by using Multi-level B-splines Approximation（MBA） algorithm to evaluate displacements of the remaining points with the set of matched feature points. Experiments show that the accuracy of projection and the effect of mosaicing panoramic images improve significantly.

Key words: street view image, projection, feature matching, image warping, panorama mosaicing

摘要： 当前主流的全景拼接软件在进行投影变换的过程中对街景影像上匹配点对的精度的依赖性较强。为了增强其鲁棒性，采取在投影后的影像上获取匹配点对并将其反投影回街景影像坐标系的策略来减少对初始匹配点对的依赖。此外，针对投影后相邻图像之间的错位导致全景拼接产生缝隙的问题，采用多层B样条插值算法对相邻图像的光流进行插值来消除影像间的错位。实验结果表明，上述的算法策略提高了投影转换的精度，明显改善了全景拼接的效果。

关键词: 街景影像, 图像投影, 特征点匹配, 图像变换, 全景拼接