A method and system for human behavior recognition based on graph convolutional network

Abstract

The invention discloses a method and system for recognizing human behavior based on graph convolution network. The recognizing method includes: extracting human skeleton information from an image containing human behavior, obtaining a sequence of human body joint point position information, and constructing an arbitrary-length topology of the human skeleton Graph sequence; feature extraction and adaptive evolution of topology structure are performed on topological graph sequence through spatiotemporal graph convolution network based on topologically learnable graph convolution, and new node features fused with local spatiotemporal features and topological graph sequence with new topology structure are obtained. ; Feature extraction through graph convolutional long short-term memory neural network; global spatiotemporal features are obtained by global pooling operation; human behavior recognition is performed based on global spatiotemporal features through classifiers. The invention directly learns the features of the entire graph, expands the weight matrix in the graph convolution to the entire topological graph structure, learns the relationship between any two nodes in the graph without being restricted by the topological structure, and has high recognition accuracy.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and relates to a method and system for recognizing human behavior based on a graph convolution network, which can be used for action recognition of topological graph sequences. Background technique [0002] Convolutional neural networks have achieved great results in many fields, but they rely on the data representation to have a grid structure. However, the data in many domains is not grid-structured, and the data in these irregular domains is usually represented as a topological graph structure, which makes it difficult for convolutional neural networks to generalize on graph domains. In recent years, some progress has been made in the promotion of convolutional structures in the graph domain. Maintaining the shared weights of convolutions between subgraphs with different numbers of adjacent nodes is an important feature of convolutional networks to achieve results in the field ...

AI Technical Summary

Problems solved by technology

The existing adaptive graph convolution can only learn the topology adaptive relationship between adjacent nodes, and has insufficient ability to learn the relationship between distant nodes.

In addition, due to the limitation of the transfer matrix in graph convolution, there is often a lack of effective modeling of long-term temporal relationships between sequences of topological graphs.

Images