关键词: CUDA, 视频拼接, 图像融合, 形态学方法

Abstract: Aiming at the problems of virtual shadow, color brightness difference on both sides of the stitching line and the inability of real-time stitching at high resolution in the multi-channel video fusion stage, this paper proposes a multi-directional perception video fusion algorithm. Firstly, in the stage of image registration, SIFT is used to extract feature points and feature descriptors to register the image. Then, in the image fusion stage, the weight lookup table constructed by exponential function is used to guide the fusion transition. Combined with the distance between the projection position of video frame and the seam, the image seam is fused adaptively by morphological operation. Finally, a large number of multi-threaded parallel operations on GPU platform are used to integrate matrix operations such as projection and fusion, so as to cover the delay and achieve the purpose of real-time. Experimental results show that the algorithm can eliminate the virtual shadow and blur in the overlapping area, and has a good real-time stitching effect.

Key words: CUDA, video stitching, image fusion, morphological method