The Study Of Motion Video Annotation Algorithm

The rapid development of computer network technology makes a convenient condition for the image,video and other information to storage and transmission.It is an urgent problem needed to solve that how to retrieval video data quickly and effectively,and video annotation is the basis and a key part of video retrieval.Motion video has a great application value because it has a wide range of market applications.In this paper,the human motion video is as the research object to study of video annotation method,in order to establish foundation for image retrieval efficiently by video annotation.This paper uses annotation algorithm based on classification,makes the annotation words as category labels,and the problem of video annotation can be regarded as a classification problem.Based on this idea,we present a motion video annotation algorithm based on multi-class SVM.The two most important factors of video annotation based on classification are effective feature extraction and representation,and the selection of appropriate classification algorithms.Most of the current motion video researches are carried against the outline of the human body,so that it can not accurately describe a variety of human motion.This paper starts from the motion analysis of human body in motion video.It is mainly divided into three parts to realize video annotation,video key frame extraction,human skeleton feature extraction and expression,video classification based on multi-class SVM.This paper adopts key frame extraction algorithm based on clustering method to extract the key frames efficiently,which can greatly reduce the redundant information of video.Make human upper body pose estimation,and then we present a method by using the human skeleton feature matrix to describe the movement of the human body,which can represent the characteristics of the human body vividly and accurately.We present a motion video annotation algorithm based on multi-class SVM.The multi-class SVM algorithm is used to train the human skeleton feature matrix,and then use the trained classifiers to annotate the corresponding motion video.Video data sets are obtained by taking pictures of people around and on the Internet.Experiments were performed using cross validation methods to verify the effectiveness of the proposed algorithm.At the same time,in order to prove the effectiveness of the proposed algorithm in image retrieval,we select the corresponding motion images in Buffy database,and make a contrast experiment between the method used in the paper10 and the algorithm we proposed,proving the effectiveness of our algorithm in image retrieval.