Skeleton data behavior identification method based on graph convolutional neural network

Abstract

The invention discloses a skeleton data behavior identification method based on a graph convolutional neural network, and the method comprises the steps: carrying out the downsampling of the time dimension of each segment of skeleton data in a training set, obtaining skeleton data with a fixed time dimension for training the neural network; dividing the input data into three channels according tothree-dimensional coordinates, conducting graph convolution with double attention on each space channel, and then combining the three channels; performing space-time convolution on the combined vectors; updating the weight of the neural network by labeling the action category; and finally, obtaining a behavior recognition model strongly related to the specified labeling result. According to the method, the topological structure and the space-time relationship of skeleton data are fully utilized, and the behavior recognition performance is improved.

Description

technical field [0001] The invention relates to the technical fields of deep learning and behavior recognition, in particular to a behavior recognition method of skeleton data based on graph convolutional neural network. Background technique [0002] Behavior recognition has always been a hot spot in the field of computer vision. In recent years, with the integration and development of deep learning and computer vision technology, behavior recognition has been widely used in video analysis, intelligent monitoring, human-computer interaction, enhanced display and other fields. The traditional action recognition method based on color video data requires a large amount of data, and it is difficult to learn the key information of the human body in the video. However, the three-dimensional human action recognition based on the skeleton is due to its high level of representation of human motion and its ability to view angles, appearance, Robustness to scale and environmental distu...

AI Technical Summary

Problems solved by technology

Among them, manual feature extraction includes the Lie group method based on human skeleton modeling, etc. These methods use relative joint coordinates to describe human actions, ignoring the absolute motion of the skeleton joints and the time information of the motion. These manually extracted features are difficult to characterize from the skeleton structure. complex non-linear mappings to action categories such as

The methods based on deep learning generally use recurrent neural network, long short-term memory network, gated recurrent unit network, convolutional neural network and other methods. These methods have achieved good results in skeleton-based human behavior recognition, but sequence-based The deep learning method is limited by the method itself, which is difficult to build a deep network and cannot learn the topological relationship of the skeleton itself, and the convolution-based deep learning method is difficult to learn the time information of the sequence

Images