A parallel method for aerial image stitching using ORB feature points

Aviation image stitching requires real-time processing, while traditional key point descriptor generating a floating-point feature vector, thus the processing efficiency in embedded hardware platform such as DSP, FPGA is not satisfactory. Recently ORB feature point is proposed, which has binary vector for feature descriptor, it greatly speeds up processing procedure for feature extraction and matching. In order to calculate the homography matrix between stitching sequence robustly, best of 2nd nearest matcher, cross-validation and RANSAC estimation are adopted. Registered images still have some color deviation in the same pixel position. If the traditional α-blending method is employed without consideration of the position information on the edge of the image, the stitching artifacts at image edges which largely affect the visual effects will be produced. A position-weighted image fusion algorithm which takes the location information of image pixels into consideration is also presented in this paper, so the image can be naturally stitched and the problem of artifacts is solved. The proposed algorithm is insensitive to complex noise presented in input image data. Furthermore, we propose a novel parallel framework for image stitching based on recent proposed ORB feature descriptor which is realized on a multicore DSP platform: TMS320C6678. With the implement of the parallel design, computing speed is obviously improved and real-time image stitching for airborne embedded application is realizable.