An Action Recognition Method Based on Graph Convolution and Capsule Neural Network

Abstract

The present invention proposes a behavior recognition method based on graph convolution and capsule neural network. The present invention obtains the spatial coordinates of human body joint points in each frame of continuous human motion images by manual marking, and further constructs the spatial coordinate vectors of human body joint points; maps the spatial coordinate vectors into high-dimensional feature vectors through multi-layer perceptrons, and combines the action association The joint point adjacency matrix is ​​constructed according to the principle; the velocity space vector of the joint point is constructed according to the space coordinates, and the acceleration space vector of the joint point is further constructed; the convolutional neural network is used to extract features, and the capsule neural network is used for action classification. The capsule convolutional neural network is constructed by connecting the product neural network and the capsule neural network in series; the trained capsule convolutional neural network is obtained by repeating the training set for multiple generations. The invention conforms to the characteristics of actual motion, and the propagation of the features on the graph is more in line with the actual situation, and can effectively retain the features for classification and improve the recognition ability of the model.

Description

technical field [0001] The invention belongs to the research category of behavior recognition, and in particular relates to a behavior recognition method based on graph convolution and capsule neural network. Background technique [0002] Behavior recognition is an extremely challenging direction in the field of computer vision. It has important applications in many aspects, such as intelligent visual monitoring. In some places with high security requirements, such as supermarkets and banks, it can detect whether there is any suspicious behavior of people in real time. . Human-computer interaction, in the field of advanced user interfaces, it is hoped that future robots can understand human behavior and respond accordingly in order to communicate with humans, which can greatly improve human living standards. Based on content retrieval, in the era of information explosion, if you can label videos, it will help you accurately retrieve the desired videos in massive data. In s...

AI Technical Summary

Problems solved by technology

[0008] Capsule neural network: The traditional convolutional neural network has an important problem. The orientation of components and the relative relationship in space are not important to it. CNN only cares about whether there are features, and the pooling layer in CNN discards A lot of information, such as important location information; CNN can only recognize an object after feeding a large amount of data, and it has not reached the real sense of inference

Images