Action structure self-attention graph convolutional network for action recognition

Abstract

The invention describes a method, apparatus, and non-transitory computer-readable storage medium that use a graph convolutional network (GCN) to identify human body actions. The method includes obtaining a plurality of articulation point gestures by the apparatus. The apparatus includes a memory storing instructions and a processor in communication with the memory. The method also includes normalizing, by the apparatus, the plurality of articulation point gestures to obtain a plurality of normalized articulation point gestures; extracting, by the apparatus, a plurality of coarse features fromthe plurality of normalized articulation point gestures using an improved spatiotemporal graph convolutional network (ST-GCN); reducing the feature dimensions of the plurality of coarse features through the apparatus to obtain a plurality of dimension-reduced features; optimizing the plurality of dimension-reduced features based on a self-attention model through the apparatus to obtain a pluralityof optimized features; and identifying the human body action through the apparatus according to the plurality of optimized characteristics.

Description

technical field [0001] The present invention relates to a graph convolutional network (graph convolutional network, GCN) for human action recognition, in particular to an improved spatio-temporal graph convolutional network with a self-attention model. Background technique [0002] Human action recognition has been actively developed in recent years due to its important role in video understanding. Generally, human actions can be recognized from multiple modalities such as appearance, depth, optical flow, body, etc. Among these modalities, the dynamic human skeleton often conveys important information and complements the other modalities. However, traditional skeletal modeling methods usually rely on hand-crafted features or traversal rules, thus resulting in limited expressive power and difficult generalization and / or application. [0003] There are many problems and difficulties in the existing methods for recognizing human actions through skeletal modeling, such as but ...

AI Technical Summary

Problems solved by technology

However, traditional skeletal modeling methods usually rely on hand-crafted features or traversal rules, thus resulting in limited expressive power and difficult generalization and / or application

[0003] There are many problems and difficulties in the existing methods of recognizing human actions through skeletal modeling, such as but not limited to low recognition efficiency, slow recognition speed and / or low recognition accuracy

Images