Human skeleton action recognition method based on difference graph convolutional neural network

Abstract

The invention relates to a human skeleton action recognition method based on a difference graph convolutional neural network. The method comprises the following steps: S1, preprocessing skeleton data; S2, preliminarily designing a difference graph convolutional neural network architecture, combining graph convolution and convolution, adopting a difference learning mode, and taking errors of different layers of graph convolution as input feedforward of continuous time frames; S3, preliminarily selecting training parameters, and conducting back propagation on the errors; S4, carrying out cross-view and cross-object training and testing on a skeleton data set; S5, finely adjusting training parameters according to test precision, and repeating the steps S3 and S4 so as to obtain training parameters with high precision; and S6, fixing the training parameters, finely adjusting the network architecture so as to obtain network architecture parameters with high precision, and carrying out human skeleton action recognition. Compared with the prior art, when an NTU data set is tested, CV precision and CS precision are both improved, the number of the parameters is small, the total number of the parameters is smaller than 1 M, and rapid and accurate recognition effect is achieved.

Description

technical field [0001] The invention relates to the fields of robot learning and computer vision, in particular to a human skeleton action recognition method based on a difference graph convolutional neural network. Background technique [0002] As a branch of computer vision, action recognition has a wide range of applications in video surveillance, human-computer interaction, intelligent driving and other fields. Through the recognition and classification of actions, the purposes of video monitoring, understanding and action prediction can be achieved. In the field of video security monitoring, it can judge whether the production operation conforms to the safety production standard according to the human body movement; in the field of human-computer interaction, it can predict the next moment action according to the current human body movement, and provide input for the action decision-making of the robot; in the field of intelligent driving , smart cars can adopt differe...

AI Technical Summary

Problems solved by technology

The existing action recognition algorithms are divided into two categories: traditional manual feature extraction and deep learning algorithms. Manual feature extraction algorithms include space-time interest point method, motion histogram method, and dense trajectory method. The stickiness is weak, and it is greatly disturbed by light, occlusion, background change, etc. The deep learning algorithm is divided into CNN-based and RNN-based algorithms, among which the CNN-based algorithm is smaller than the RNN network model, and the calculation memory occupies less. It is a two-stream convolution method and a 3D convolution method. The graph convolution algorithm is a popular algorithm in recent years. It has the advantage of strong learning ability, but its calculation amount is relatively large, and it consumes more memory.

Images