Human body behavior recognition method based on multi-stream three-dimensional adaptive graph convolution

Abstract

The invention discloses a human body behavior recognition method based on multi-stream three-dimensional adaptive graph convolution, and the method comprises the steps: 1) collecting a video containing a human body, and constructing a training set; 2) constructing an adaptive spatial domain attention matrix SAM and an adaptive time domain attention matrix TAM by using the training set; 3) constructing an adaptive spatial domain attention map convolution module ASAGCM, an adaptive time domain attention map convolution module ATAGCM, and a three-dimensional time-space domain map convolution module GCN-3d; 4) constructing an adaptive graph convolutional layer; 5) constructing an adaptive graph convolutional network by using the adaptive graph convolutional layer; 6) constructing a multi-stream three-dimensional adaptive graph convolutional network by using the adaptive graph convolutional network; 7) training the multi-stream three-dimensional adaptive graph convolutional network by using the training set; and 8) utilizing the trained multi-stream three-dimensional adaptive graph convolutional network to perform behavior recognition on the human body in the video. According to the invention, the accuracy of human behavior type identification can be effectively improved, and a good foundation is laid for various computer vision processing applications.

Description

technical field [0001] The present invention relates to the technical field of image pattern recognition and deep neural network, in particular to a human behavior recognition method based on multi-stream three-dimensional adaptive graph convolution. Background technique [0002] At present, human behavior recognition in video is one of the most active research topics in the field of computer vision. It has broad application prospects and potential economic benefits in intelligent video surveillance, human-computer interaction, content-based video retrieval, virtual reality, etc. value. [0003] Traditional behavior recognition algorithms mostly use RGB video as input, but when the RGB video has a lot of background dynamic interference, unstable lighting, and severe noise, the effect of traditional behavior recognition algorithms will be affected. In recent years, compared with traditional methods using RGB videos for recognition, skeleton-based action recognition has attra...

AI Technical Summary

Problems solved by technology

The skeleton connection input by the existing skeleton-based behavior recognition algorithm is defined according to the natural connection of the human body, which does not make full use of the information between the skeleton joints, and does not make full use of the front and rear frame information of the joints. The utilization is relatively independent, and the spatiotemporal information between the joints is not fully effectively combined. Therefore, it is urgent to propose an adaptive graph convolution recognition method that adaptively adjusts the skeleton connection and fully integrates the spatial and temporal information.

Images