Research On Video Stitching Keyframe Extraction Technology Based On Feature Statistics

With the development of digital technology and multimedia technology,the increasing demand for videos capture is increasing in their work and life.A varity of cameras are gradually increasing.The popularity of panoramic cameras is not high,and the photos taken by a single camera sometimes cannot meet people’s requirements for wide viewing angles.How to splice videos under multiple shots and effectively process video information has become an important problem to be solved urgently.Therefore,this paper aims to study a high-quality video stitching method,and extract the key frames in the video after the stitching is completed,so as to quickly obtain the main information in the panoramic video.It is expected that it can be used in monitoring,automatic driving and other fields in the future.This topic collects pictures and videos from ukbench and Open Video Project datasets for research,and uses image noise reduction,image enhancement and other methods to preprocess the collected pictures.Based on the commonly used key frame extraction method and the basic steps of image stitching,the experimental results show that the algorithm with better stitching quality often runs slower,which does not meet the real-time requirements of video stitching.For different types of videos,there are often a large number of redundant frames in the extracted key frames,which does not meet the requirement that key frames can accurately describe the video content.This thesis includes two aspects,optimizes and improves the video splicing speed and the quality of key frame extraction.The main contents are as follows:Firstly,improve the SURF algorithm.In the feature point description stage,the area farthest from the center point is eliminated,and at the same time,only the feature points in the overlapping area of the two images are extracted.Experiments show that the improved algorithm has the advantages of being fast and efficient without affecting the splicing quality.Secondly,optimize video frame-by-frame splicing.Compare the frame difference between two adjacent frames.Compared with the previous frame,the frame with obvious feature point changes is defined as a dynamic frame,otherwise it is defined as a static frame.First save the homography matrix of the first frame,the static frame directly uses the saved homography matrix,and the dynamic frame recalculates the homography matrix and saves it.Experiments show that this method has a good splicing effect for both static and dynamic videos.Since the algorithm reduces the steps of repeatedly calculating the homography matrix of the static frame,the running speed of the algorithm is improved.Finally,condense the keyframes.Improve the traditional frame difference algorithm,select the frame corresponding to the local maximum point as the candidate key frame,and use the adaptive threshold to carry out one-layer clustering on this basis,in order to meet the needs of different users,use the manual setting threshold for the second-layer Clustering,select the frame with the largest average mutual information with other frames in each cluster after two-level clustering as the key frame.Experiments show that the number of long video key frames can be reduced by about 70% after clustering.