Research On Human Body Action Recognition Method Based On Video Feature

For video sequence of human actions, its identification and classification in the field of computer vision and in civilian areas of life has a wide range of needs. Intelligent traffic monitoring and neighborhoods have been behind the people’s demands, in urgent need of upgrading surveillance systems to monitor efforts to strengthen security in residential areas as well as to improve the detection accuracy of the traffic violation. Improving the recognition rate of unlawful actions can increase detection rate of public security, also has a great contribution on the the maintenance of social order.But now on human action recognition technology is not mature, basically still more strongly dependence. on the computer, less on the people. Therefore,there are great value and practical significance in human motion analysis and research.In this paper, we mainly study the three simple body actions:run, walk, jump, include these simple actions’ detection, classification and identification is in a single scene, based on the Weizmann database’s videos. The research process is divided into body moving target detection, feature extraction, and human action recognition,presenting these three steps into the trend, a step can not be missing.First, the detection results of the human motion have a direct impact on action recognition rate. the paper cited the more commonly used two methods:inter-frame difference, background subtraction method. Introduce their basic principles and the advantages and disadvantages of the two methods, and compare the effect of these two detection methods. Based on experimental requirements, decide to adopt background subtraction method to extract the target area. Get the body edge contour with pretreatment.Secondly, extract the body’s feature according to the obtained edge contour. The body feature extraction method is mainly studying the PCA method, grid method, focus on the human skeleton method. Law on human use of human skeleton edge contour feature extraction, followed by the use of rotating from its serialized, the final analysis of the results of the match and the matching accuracy.Finally, according to the sample matching the observed sequence as serialization, use Hidden Markov Model for human action recognition.Extract the feature of different people’s actions as the test samples, match the code table created by the feature of the training samples to get sample sequence. Using Baum-Welch algorithm to estimate the model parameters constant, so that get the output maximum of the probability of the sequence. Also say the highest matching rate to reach the recognition results.The result of the experiment show that it is successful to combine the feature extraction based on human skeleton and human action recognition based on HMM, because of the better recognition results. It also has good versatility and robustness.