Method for constructing human body behavior recognition model based on graph convolution network

Abstract

The invention discloses a method for constructing a human body behavior recognition model based on a graph convolution network. The method comprises the following steps: 1, acquiring and preprocessinga skeleton sequence; 2, constructing a time-space diagram representing the skeleton sequence; 3, constructing a three-flow graph convolutional network based on the space-time graph; wherein the three-flow graph convolutional network comprises three networks used for modeling three kinds of information including joint points, bones and bone movement of an input space-time graph respectively, and three corresponding graph convolutional networks, and outputs of the graph convolutional networks are fused to serve as outputs of the three-flow graph convolutional network; and 4, converting the skeleton sequence obtained in the step 1 into a space-time diagram through the step 2, and inputting the space-time diagram into a three-flow diagram convolutional network for training to obtain a human body behavior recognition model. According to the method, three kinds of information of joint points, bones and bone movement are calculated through the space-time diagram and are used for training themodel, so that action recognition is carried out by using richer skeleton information, and the recognition performance is remarkably improved.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a method for constructing a human action recognition model based on a graph convolution network. Background technique [0002] Action recognition tasks have applications in many aspects. Among them, human skeleton information has many advantages, such as strong robustness to scale changes, background interference, etc., and skeleton information is easy to pass through hardware devices (kinetics) or software (openpose). )get. The traditional skeleton-based action recognition usually designs features manually to represent the skeleton. There are currently three deep learning methods to solve the skeleton-based action recognition problem. The joint point information is represented as a pseudo image, and then predicted by CNN; the joint point information is represented as a graph structure, and the graph convolution is used for prediction. The first two methods cannot extr...

AI Technical Summary

Problems solved by technology

However, the topology of the graph applied in ST-GCN is fixed on all layers, lacking the flexibility and ability to model the multi-level semantic information contained in all layers;

[0007] (3) A fixed graph structure may not be optimal for all samples of different behavior categories

This fact suggests that the graph structure should depend on the data, however, ST-GCN does not support this dependency

Images