Research On Human Motion Recognition Based On Skeleton 3D Information

Human behavior recognition refers to the automatic perception of human state and action intention by establishing the relationship between low-level data and high-level semantics from images or videos and other data.Learning differentiated spatio-temporal features from bone sequences is a key factor for behavior recognition.How to analyze human motion information and understand its temporal characteristics is a challenging problem,which integrates motion detection,target classification and tracking,understanding and recognition of human motion and other multidisciplinary knowledge.Thanks to the progress of artificial intelligence and computing power,human motion recognition technology has been integrated into many fields of our daily life,such as surveillance,motion analysis,human-computer interaction and so on,and has achieved fruitful research results.In this paper,the three-dimensional coordinates of the skeleton are used as the original data,and the human skeleton is abstracted into two elements--joints and bones.Where,nodes correspond to points in the figure,and bones correspond to edges in the figure.Thus,the skeleton can be simplified as a sparse graph composed of points and edges,and the space-time joint relationship can be modeled with the graph.In three-dimensional Euclidean space,the properties of the points are(x,y,z),and the edges are line segments in three-dimensional space.In addition,time dimension is added in this paper to form skeleton sequence.The continuous changes of human posture within a certain period of time are defined as actions,such as jumping,waving,sitting,etc.The three-dimensional skeleton data uses a set of three-dimensional coordinates to represent the human body structure,avoiding the influence of background noise,light changes,and appearance changes.Skelet-based motion recognition is a research hotspot in the field of computer vision,which excavates spatio-temporal characteristics of human motion from a given skeleton sequence.In order to realize the human motion recognition technology based on skeleton information,a graph convolutional network is introduced in this paper,and the skeleton graph is divided into four subgraphs,each of which shares a joint.Each subgraph corresponds to a part of the human body,and the partial-based graph convolutional network is used to learn the recognition model,capture the advanced attributes of the subgraphs,and learn the relationship between them.Learning the discernable spatiotemporal features is the key element of human behavior recognition.In order to further improve the accuracy of recognition,geometric features and motion features are used to replace three-dimensional joint coordinates as node features on the basis of skeleton spatiotemporal map to improve the model performance.In order to improve the accuracy of the Network,an Attention Enhanced part-based Graph Convolutional Network(AEPB-GCN)is proposed.The algorithm takes the improved geometric features and motion features from the three-dimensional coordinates of the bones as input,and uses the graphic convolutional neural network to extract the features while integrating the attention mechanism.The purpose is to increase the importance of the important channels,so as to improve the accuracy of the network.The experimental data in this paper are mainly NTURGB+D dataset and HDM05 dataset.The former uses cross-subject(CS)and cross-view(CV)to evaluate the accuracy of the algorithm,while the latter uses ten-folding cross-validation to evaluate the accuracy of the algorithm.The experimental results show that the experimental accuracy of the algorithm in this paper has been improved to a certain extent.Good results have been achieved.