Cross-scene human body action recognition method based on adversarial meta-learning

Abstract

The invention discloses a cross-scene human body action recognition method based on adversarial meta-learning, which comprises the following steps of: firstly, acquiring a human body action CSI (Channel State Information) signal by using wireless signal receiving and transmitting equipment, then processing an original signal by adopting a discrete wavelet filtering technology, obtaining a CSI sample by using a threshold segmentation method, and constructing a meta-learning task set; sequentially training a feature extractor module, a generator module, a discriminator module and a human body action recognition module by using the task set to obtain a basic adversarial meta-model; and finally, collecting a small amount of data in a new scene to finely adjust model parameters, so that the model parameters can better adapt to recognition of new action types. According to the invention, high recognition accuracy of new action types can be obtained with low cost when scenes, users and the like change.

Description

technical field [0001] The invention relates to the fields of wireless network, human action recognition and deep learning, in particular to an adversarial meta-learning method based on few samples. Background technique [0002] In recent years, with the development of Internet of Things technology, human motion recognition system based on radio technology has been widely used in smart home, health monitoring and other fields, so it has attracted more and more attention. Existing methods can be divided into WIFI, ultra-wideband (UWB) and radio frequency identification tags (RFID) and so on. Among them, UWB technology requires the deployment of a large number of devices to receive and process signals, and the need to modify the site, resulting in high costs; and the realization of human body recognition based on RFID tags, mobile phones or wearable devices also requires the target to carry equipment. At present, the WIFI-based human motion recognition system is becoming a po...

AI Technical Summary

Problems solved by technology

However, this method is very impractical, because this calibration process will cost a lot of manpower and material resources

Some plants deploy fixed hardware to get new CSI for modification, but additional hardware implementation incurs additional cost

In addition, another challenge to consider is that when a human recognition system trained in a known environment is deployed to a new scene to recognize the actions of a new user, the recognition accuracy will drop significantly

This is because people usually repeat the same action at different speeds in different environments, which leads to excessive volatility of the same CSI action signal, which cannot be recognized by the system

The system must be retrained for different scenarios, which will bring a considerable amount of repetitive work to action recognition

Images