Research And Implementation Of Human Action Recognition Baesd On HMM

The human action recognition is a hot topic in the field of virtual reality，pattern recognition and artificial. In addition, it has widely used in human-computerinteraction, athletic performance analysis, intelligent video surveillance system, etc.Explaining the monitored events that occurred in the scene is the main task of actionrecognition and the four aspects of the content are moving object extraction, target tracking,feature extraction and human action recognition.The study and analysis of four aspects are the main research of this paper. And on thisbasis we focus on the feature extraction and human action recognition. The main work of thispaper as follows:In the image preprocessing, using background subtraction methods silhouette map ofhuman behavior, and then morphological processing, improve the quality of the image, toobtain satisfactory images of human movement. Finally, based on the Kalman filter algorithmCamshift to realize human tracking.In the feature extraction, single feature of the problem cannot be good description ofhuman action, using motion features and Zernike moments. The paper using the trajectoryfeature and so on as motion features the outline response characteristics of human action;using the improve posture Zernike moments feature characterize the physical characteristicsof the relative position of other details.In the human action recognition, the paper extracts thirty improved Zernike momentfeatures and the motion features can be calculated as human motion characteristics; forcreating the best model to describe the action and accurately identify action, we use thedouble hierarchy of HMM which can express the internal connection between two features toclassify the action. Using the model, the number of states to determine the most suitablemodel based on the number of body movements critical stance.Experiments are carried out in common standard video datasets, experimental results show that the combined features of the paper has better stability and representativeness, canbe fully extracted from the global and local action information and extract the information ofhuman action in the image sequence completely, Robustness and the recognition performanceis better than that of single feature and other combined features. The FDH-HMM is moresuitable for human motion characteristics, the recognition result is better than other methods.