CUDA并行计算下基于扩展SURF的多摄像机视频融合方法

Multi-video fusion with extended SURF based on CUDA parallel computing framework

在多摄像机视频融合过程中,需要对多个摄像机获取的视频中的每一帧图像进行大量诸如特征提取、图像配准、图像融合等高复杂度的计算,占用大量的运算时间,这对视频融合的实时性要求是一个很大的挑战.基于CUDA(Compute Unified Device Architecture)并行计算框架,提出了一种快速、可靠的多摄像头视频融合方法,该方法首先利用基于局部环形扩展及颜色描述子的SURF(speeded up robust features)特征提取方法提取图像特征点;其次采用基于分块相似性度量的k-d树(k-维树)多图像自动特征匹配算法进行图像与特征点的匹配;然后使用RANSAC(Random Sample Consensus)算法计算变换矩阵;最后使用多频率融合算法进行多摄像机视频融合,得到流畅的大视场视频.整个多视频融合过程使用CUDA进行并行加速,并在多个不同场景与摄像机数量下的实验验证了本文算法的实时性与有效性.

In the process of multi-camera fusion,it is necessary for each frame image acquired by multiple cameras to carry out a large number of image processing operations,such as feature extraction,image registration and image fusion.These operations take up a lot of computing time and are difficult to meet the requirements of real-time video fusion.This paper proposes a fast and reliable fusion method with multi-video camera based on CUDA（Compute Unified Device Architecture）parallel computing framework.This method firstly uses the extracting method of characteristics of SURF（speeded up robust features）which is based on partial annular extension and color descriptors to extracted image feature points;secondly,based on the block similarity measure tree,uses multi-image auto matic featurematching algorithm tomatch the images and the feature points;and then uses the RANSAC（Random Sample Consensus）algorithm to calculate the transformation matrix.Finally,multi-frequency fusion algorithm is used for multi-camera fusion.At the same time,the whole multi-video fusion process uses CUDA to carry on the parallel acceleration.The real-time and the effectiveness of the proposed algorithm are verified by experiments with a number of different scenarios and cameras.