Human body behavior recognition method based on activity graph weighting in multi-cross-domain scene

Abstract

The invention discloses a human body behavior recognition method based on activity graph weighting in a multi-cross-domain scene. The method comprises six parts of human physiological signal preprocessing, signal activity graph conversion, model pre-training, sample weighting, model training and motion state recognition. Extracting each dimension of the preprocessed signal sample, rearranging and combining, and converting into an activity graph form by applying two-dimensional discrete Fourier transform; then, on the basis of the hypothesis that data of the same category can be mapped to close positions of the subspaces, source domain data and a triple loss pre-training model are applied, and the generalization ability of the model is improved; weighting samples of a source domain and a target domain by utilizing probability prediction of a domain discriminator; source domain weighted cross entropy loss and target domain weighted information entropy loss are calculated, and the entropy of the overall prediction probability is used as a regular term to achieve the purpose of cross-domain knowledge migration; and finally, inputting target domain data into the model, and classifying human body activities through a softmax function.

Description

technical field [0001] The invention relates to the field of biomedical signal processing, in particular to a human behavior recognition method based on activity graph weighting in multi-cross-domain scenarios. Background technique [0002] Human Activity Recognition (Human Activity Recognition) has become a very attractive research area in recent years due to its potential application value, such as healthcare monitoring, fall detection, and smart home perception. The core of traditional human action recognition methods is to use enough labeled human physiological signals to train accurate and robust models. However, obtaining sufficient labeled data is time-consuming and labor-intensive. Transfer learning is widely used in human action recognition to solve label shortage and cross-domain recognition problems due to its ability to transfer knowledge from labeled source domains to unlabeled target domains. Existing human behavior recognition algorithms based on transfer le...

AI Technical Summary

Problems solved by technology

In fact, the complexity of manual features increases the difficulty of knowledge transfer, because statistics-based manual features do not necessarily contain transferable knowledge, and increase computational complexity

Therefore, these methods are only suitable for a single cross-domain scenario, and it is difficult to apply to human behavior recognition in multiple cross-domain scenarios.

How to unify the differences in data distribution in various cross-domain scenarios, design and implement multi-cross-domain scene human behavior recognition methods that do not need to extract complex manual features is a big challenge

Images