Human Action Recognition Based On Key Poses

With the development of computing power of computer central processor,disk storage capacity and various improving of video devices,now the collection and storage of video data is more and more easy, video data has been widely used in our daily life,people take the network to share their video and pictures,the video data on the network is increasing. In some of larger of people places,such as hospitals, banks, airports and train stations, video monitoring is a very important means to monitor and deter crime. But these camera and storage devices can only collect and store the video data, and a variety of actions and behavior in video human purpose still need manual interpretation because they do not have the ability of automatic recognition. The intensity of work in the video can be effectively reduced by human action recognition technology, and improve work efficiency.Because of complexity of the structure of human body, can do many actions, so this technology of human action recognition has been facing many problems and difficulties. In the case, a large of approach have been proposed for human action recognition,recognition method based on feature and region has been studied and developed deeply, but because it is collection of various video action in a period of time, so in the study of it is still face a variety of constraints, with the spatial and temporal characteristics are presented, some of the temporal and spatial characteristics of the whole dynamic system based on human action recognition has been proposed, temporal and spatial features can be very good in time and space on the basis of describing human action, has become a hot topic for human action recognition, but also has the information of temporal and spatial characteristics of larger problems. In addition, too much of the body can show the action category, so all the actions of the classification and statistics is almost impossible.In order to solve these problems of speed slow and low accuracy in traditional method, this paper proposes the use of a few key action comparative position to the recognition of human action in video, this method can improve the speed and accuracy to some extent.This paper has three parts, the first step is to extract the key posture samples and Hu moment of samples. It is through the analysis of a sequence of video frames according to different body movements, movements in the frame of self similarity and periodic motion feature, artificial chooses the appropriate key action as a sample, and then computing the Hu moments of these key actions, and to save them. The second step is to identify the video motion. To measure the video, to extract moving objects, compute the invariant moments of it, and the key poses sample moment invariant, then the K nearest neighbor algorithm with refuse to treat sample classification, finally will not in the existing classification sample categories identified as unrecognized motion, and put these unrecognized action to collect a particular collection. Finally, according to the action of the meaning of these unidentified action, modeling, and extract the key action, build a new action categories.The experimental results show that some characteristics of human action action can effectively describe the actions and distinguish between different action, selecting the key action can improve the accuracy of identification of effective, by the absence of recognition in data classification, collection and identification of these actions, it can improve the accuracy of system the classification of the query.