Grouping confusion graph convolution action recognition method based on skeleton information

Abstract

The invention discloses a grouping confusion graph convolution action recognition method based on skeleton information. According to the method, a grouping thought is adopted, when space operation is carried out, the dynamic graphs are grouped to extract information of different graph structures, and rich behavior information is obtained. Meanwhile, the parameter quantity of the model can be reduced through the grouping form. Then, when operation is carried out on a time sequence, a depth separable convolution form is adopted to reduce parameters and a calculation amount. Because a grouping form is adopted in space and time sequence, information of different groups needs to be fused so as to achieve information circulation. The result shows that under the condition that high performance is kept, the parameter quantity and the calculated quantity are greatly reduced, and the effectiveness of the method is proved.

Description

technical field [0001] The invention is a grouped confusion graph convolution action recognition model (GS-GCN) based on skeleton information. The introduction of grouping is to obtain multiple dynamic graph structures, so as to obtain some activities that cannot be represented by the original graph structure, such as clapping, brushing teeth, running, etc. Using multiple dynamic graphs can extract richer information and reduce parameters and calculations. We use graph convolution to extract spatial information, and the skeleton information itself, as a graph structure, is very suitable for operation with graph networks. The depthwise separable convolution is introduced on the timing module, and the depthwise separable convolution can maintain high performance while reducing parameters and calculations. Since we use the form of grouping in both space and time sequence, we need to integrate the groups to achieve the flow of information. Finally, we use the network to identif...

AI Technical Summary

Problems solved by technology

[0009] (2) The current method has a large amount of parameters and calculations

In order to maintain high performance, a lot of work is in the form of module stacking, generally 10 modules are stacked, which leads to a high final parameter value, which is not good for practical applications

Some mobile devices cannot run larger models due to insufficient computing power

Images