Research On Human Action Recognition Of 3D Skeleton Based On Non-Euclidean Convolution

The advancement of science and technology has brought us a better life,and the needs of real life have enabled science and technology to have a stronger development momentum.Artificial intelligence has been reflected in all aspects of life,bringing people an unprecedented interactive experience.Human action recognition is a research closely related to life and an important field of artificial intelligence.Deep analysis of the acquired complete 3D visual data of the human body is a frontier research topic in the field of machine learning and pattern recognition.Two important research directions of action recognition,one is how to extract more robust and distinguishing features,and the other is how to extract spatial and temporal feature information better from the acquired features.The learning method using RGB video will be affected by the background and lighting,and there is also the problem of occlusion.The direction of human motion and the angle of video shooting will also have an impact.Based on the RGB video method,CNN is generally used to extract spatial features,or add features extracted from each frame of image or optical flow features extracted from video images.The above factors have an inevitable impact on recognition.With the development of depth cameras,skeleton information of the human body is easier to obtain.Compared with RGB video information,skeleton information not only takes up less resources,but also has the advantage of not being affected by factors such as background.The advantages of skeleton make action recognition based on skeleton become an important research direction.In this paper,two frameworks for action recognition are proposed based on skeleton data.The original skeleton coordinates are greatly affected by the view variations.Representing the skeleton data as the rotation transformation relationship between each pair of bones can not only eliminate the influence of the view variations,but also express more information.Therefore,this paper proposes to calculate the Lie group transformation information of each pair of bones,and express the entire skeleton sequence as a 3D Lie group cube feature representation method.Around the 3D Lie group cube feature,this paper defines the convolution and pooling operations on Lie group by expanding the image convolution.According to the form of Lie group features,convolution and pooling on 2D Lie group and 3D Lie group are defined.It also proposes a Lie group convolution neural network working on Lie group space.Aiming at the Lie manifold constraint parameters existing in the learning process,the optimization of parameters in the manifold space is discussed.This work gives a new representation of the skeleton information,puts it in the Lie manifold space,gives a method of extending the convolutional network into the manifold space,and learns the manifold features through logarithmic mapping.Combining conventional deep learning layers.From the skeleton structure of the human body,a topology structure can be extracted.The second work in this paper is based on the original skeleton coordinate data and proposes a graph neural network that can automatically adapt to graph topology.This paper discusses the difference of contribution of different neighborhood of nodes to nodes from multiple perspectives,puts forward the method of dividing the neighborhood of nodes into different regions and learning the convolution of the graph respectively in the directed graph,and discusses the difference between the directed graph and the undirected graph.According to the different importance of different neighbors to the motion,the spatial attention mechanism of the graph is discussed,and the form and effect of the attention mechanism on the adjacency matrix are discussed.Aiming at the structure of the original graph that is too fixed to modify the adjacent edges,an adaptive graph convolution scheme is proposed.The main contribution of this work is to propose a graph convolution method on directed graphs and an adaptive graph network.The test and evaluation of the proposed method on multiple public datasets proves the effectiveness and advanced of the proposed method for action recognition.