| The video splicing technology converts the video sequence image into a largefield-of-view and wide-angle panoramic image that contains all the information of the sequence,and solves the problem of the small shooting angle and the inability to observe the panorama due to limitations of the camera equipment.Panorama video splicing technology has been widely used in virtual reality,security monitoring,auxiliary vehicle,remote desktop and the like.Video splicing technology is considered as an important branch of computer graphics and computers.Its development is based on the splicing technology of static images.Image splicing focuses on splicing effects,and video splicing also needs to meet the requirements of real-time.The splicing technology is mainly divided into registration and fusion stages.This paper mainly studies the registration phase in feature-based video splicing.In the video registration phase,traditional image-based registration algorithms are used,and the search area is often designated as an entire image,resulting in a large search area,too many feature points,and a large number of invalid feature points and mismatches,particularly high The resolution image can not meet the real-time requirements of the video.At the same time,the RANSAC algorithm used for the purifying is used.In some scenes,the geometric characteristics of the matching point pair are ignored and random sampling strategies are adopted,resulting in an excessively high number of iterations.For the problems in the above registration phase,the main contents of this paper foucus on the following four aspects:(1)A quick registration algorithm based on overlapping regions is proposed.For the parallel and aggregating camera setup methods,the camera's position and field of view are geometrically modeled.Based on the ratio of the overlapping area to the camera's field of view,the ratio of the blended area image to the camera image is given,and the feature point is reduced.The search area reduces the number of useless feature points and speeds up the registration process.(2)An improved RANSAC algorithm is proposed.According to the camera setting method,the geometrical characteristics of the registration point's slope are considered;the quality evaluation criteria of feature points are defined to further refine the matching point pairs;in order to ensure the robustness of the algorithm,a slope grouping strategy and a decision rule are introduced.(3)Accelerate the registration process using GPU parallelization technology.According to the characteristics of the CUDA programming model,the GPU parallelization analysis and implementation of the SIFT feature extraction and matching are performed to speed up the calculation process of the Gaussian pyramid construction,local extremum detection,feature point direction gradient,feature descriptor,and RANSAC purification.(4)Designed and implemented a video splicing system.By calculating the parameters in the off-line phase and completing the integration design in the online phase,the entire splicing system is divided into three modules: a pre-processing module,a registration module,and a fusion module.In the experiment,on the camera with a resolution of 1920x1680,the computational effects of the fast registration algorithm and acceleration strategy studied in this paper on the entire video mosaic are tested. |
