基于CUDA的SIFT特征与拼接缝的全景图生成

Panorama Generation of SIFT and Stitch Line Based on CUDA

针对大场景高清全景图的生成需求,提出了一种基于CUDA的SIFT特征与动态规划拼接缝的全景图生成方法。首先进行柱面投影保持各通道视频在重叠区域内成像的视觉一致性,并设计了基于CUDA的SIFT特征提取并行算法和基于RANSAC与特征点的变换模型鲁棒估计算法;然后通过基于HIS色彩空间与梯度空间的动态规划拼接缝实现相邻图像的缝合,并在缝合线两边采用基于三角函数的图像融合算法实现平滑过渡,最终生成高清晰度高分辨率的全景图。结果表明,该方法在满足图像无缝拼接精度的前提下,可消除重叠区域内的重影问题,并大幅度提高拼接速度。

Aiming at the effect of high-definition panorama generation of large scenes,propose a panorama generation method based on dynamic programming seam and SIFT based on CUDA. Firstly,the visual consistency of the overlap regions is guaranteed by the cylindri-cal projection,also design a parallel algorithm of SIFT feature extraction based on CUDA and a robust parameter estimation algorithm based on RANSAC and feature points. Then,the adjacent images are stitched by dynamic programming stitch line based on HIS color space and gradient space. A trigonometric-based image fusion algorithm is used on both sides of the stitch line to realize the smooth tran-sition. The high-definition and high-resolution panorama is finally generated. The proposed method can eliminate the ghosting problem in the overlapping area and significantly improve the stitching speed,on the premise of high-accuracy seamless image mosaic.