Research On Video Retrieval Method Based On Human Pose

Video retrieval has important applications in the field of video surveillance and educational video,and it is a very hot research direction at present.In recent years,most of the research on video retrieval technology is based on content.Now,in the existing video media,a large part of the video content is related to the daily behavior of human beings,very few studies have analyzed human motion pose and used the estimated pose to retrieve the video,the main reasons are the cluttered background in the video,the diversity of human pose,the variety of actions and the self-occlusion of human body,etc.Aiming at the problem of video retrieval based on human pose,this paper emphasis on reducing the part search space and human pose estimation in video,and the key frame extraction method based on human pose,and have finished the the retrieval of video containing human daily action.Firstly,an improved SLIC algorithm combined with GrabCut algorithm is given to reduce part search space and to estimate the human pose by using interframe motion information.When using the SLIC algorithm to generate superpixels,we do not need to specify the number of superpixels expected,but rather specify the number of pixels that each superpixel contains,divide the image into areas of the same size and select one or two superpixel centers in each area.The GrabCut algorithm extracts the foreground with the unit of super pixels generated by the improved SLIC algorithm,human body part only needs to be searched in superpixels,the motion information between frames is used to generate the better candidate poses for a single frame image,and the sequence of body parts is reorganized to find the optimal pose.Then,a key frame extraction method based on pose is proposed.The difference of human body changes in different movements is mainly reflected in the change of limbs.The joint part of human body is the key area of action,and the change in the joint angle of the lower extremity contains more abundant information.Firstly,the angle sequence of lower extremity joint is determined,and the dense optical flow is calculated for each region to form the pose motion information.Then,by calculatingthe differences of angle sequence and motion characteristics between frames,and comparing the linear combination of differences with the set threshold to determine the choice of key frames.Finally,the key frame is described by extracting the human pose feature and forming a feature descriptor based on pose.Then matching with the pose feature descriptor in the database to achieve video retrieval.The video content in the data set is the daily action of the human body.The experimental results show that the human pose estimation method can improve the speed and accuracy of pose estimation,the key frame extraction method based on pose can highlight the main contents of video,control redundant frames and retrieve video effectively.