Specific Motion Recognition Based On Human Body 3D Skeleton Model

In recent decades,there have been terrorist attacks and violent incidents in public places,which seriously threaten the people's lives.Intelligent video surveillance technology has developed rapidly in recent years.It is an effective way to prevent the occurrence of terrorist violence by arranging a dense camera in public places with high traffic flow,automatically identifying pedestrian dangerous actions and issuing early warning signals to remind security personnel to take timely measures.This paper studies the recognition method of intelligent video monitoring technology for specific action.The standard human skeleton model was established by selecting 15 important joints in the human body,and the similarity study was carried out based on the skeleton model to judge whether the action belonged to the specific action.The main work of this paper has the following points:(1)In recent years,the appearance of inexpensive depth sensors has made it easy for researchers to obtain RGB-D images to extract the human skeleton information.Compared with traditional RGB images,RGB-D images are less affected by light and background changes.In this paper,the features extracted in action recognition are stratified according to time and spatial hierarchical relationships,and are divided into seven levels.The high level features can be represented by the low level features.This classification method facilitates the study and understanding of the feature extraction methods of the human skeleton model.In fact,the bottleneck features represent the joint coordinates and other depth image data,while high-level features represent actions,the selected feature is like a bridge of low-level features to high-level features.(2)A feature vector matching algorithm based on the main features is proposed.The traditional eigenvector matching algorithm matches all the eigenvalues of the extracted eigenvalues,but does not take into account the fact that each feature is different in importance.The improved method of this paper draws on the idea of selecting the main components in PCA principal component analysis,calculates the variation range for each eigenvalue in each action sequence,and selects several features with the largest variation to form a new vector to match.(3)A method of constructing a decision tree based on local limb movements is proposed,and the difference of local limbs between different actions is taken as the decision condition of decision tree classification.Finally,the two open data sets are used to test the accuracy of the two action classification methods.