Research On Video Stitching Algorithm Based On Feature Matching

As an important technology in the field of computer vision,video stitching technology has been applied widely used in virtual reality,panoramic photography,medical images and other fields.The video stitching technology is a technique of combining a picture of a certain overlapping area and a picture with a certain overlapping area into a wider picture field,which has higher requirements for better splicing effect and real-time performance.At present,there are still many problems in the implementation process of the video stitching technology based on traditional algorithm method : the real-time performance difference of the algorithm affects the video fluency,and the robustness of the algorithm leads to the existence of heavy lines and broken lines in the video picture.This paper has done some research work over the above issues,the main contents are as below:1.This paper firstly analyzes the research background and research status of video stitching,and summarises the key problems that need to be solved in the current video splicing technology.Then it introduces the main process of image stitching,image transformation model,image registration technology,image interpolation technology and image fusion technology.In the end,the advantages of video stitching based on feature matching are expounded through comparative analysis.2.This paper analyzes and compares the classical feature algorithms,and proposes a feature matching algorithm based on scale FAST and improved LBP descriptor according to the characteristics of video stitching.At the stage of detecting feature points,a scale pyramid is constructed and a very fast FAST algorithm is used to extract feature points to enhance the robustness of scale invariance.Then,the improved LBP description method is used to describe the feature points,which enhances the robustness of rotation invariance and improves the matching efficiency by reducing the dimension of feature vectors.Finally,DDRN algorithm is used to measure the similarity of feature vectors to complete the matching,and the improved RANSAC algorithm was used to eliminate mismatches.Through experimental verification,compared with the traditional SIFT and SURF algorithms,the algorithm proposed in this paper not only greatly improves the real-time performance,but also accurately achieves the feature extraction and matching of the images.In addition,the improved LBP description method brings a strong anti-interference effect on the rotated image,and is still highly adaptable to complex transformation scenes such asaffine,scaling and illumination.3.In this paper,the commonly used fusion algorithms are introduced,analyzed and compared with experiments,the basis of selecting multi-resolution fusion method based on Laplace pyramid is expounded,and the effectiveness and real-time performance of this algorithm for different image fusion scenarios are verified.Aimed at the requirement of high real-time performance in video stitching,a video inter-frame processing method based on inter-frame difference is proposed.By setting the threshold to compare the values of difference between frames,and judge the distinction between static and dynamic frames.Static frames directly use the saved homography matrix,and dynamic frames recalculate the transformation matrix.By setting up the experiment environment of video stitching,the static video and dynamic video are stitched and verified separately.The results show that the proposed method achieves better splicing effect for both static and dynamic video,and the algorithm reduces the repeated calculation of static frames,so the running time has been greatly reduced,meeting the requirements of real-time video stitching.