Incremental learning human body action recognition method based on CSI (Channel State Information)

Abstract

The invention discloses a CSI (Channel State Information)-based incremental learning human body action recognition method. The method comprises the following steps: 1, collecting CSI action samples in M + 1 scenes by deploying WIFI (Wireless Fidelity) equipment; 2, adding time information to a CSI action sample; 3, generating a pseudo sample for each type of action by using a data enhancement method; 4, finishing cross-scene human body action recognition by utilizing an incremental learning model; and 5, applying the updated model and the memory set to the learning of the next new scene action. According to the method, human body actions in different scenes can be continuously learned, and the requirement on the quantity of action samples is reduced on the premise of ensuring relatively high recognition precision.

Description

technical field [0001] The invention belongs to the technical field of wireless communication, in particular to a CSI-based incremental learning human action recognition method. Background technique [0002] Human motion recognition has been widely used in smart home, gesture recognition, daily behavior detection and other fields. Compared with computer vision perception technology, infrared perception technology and special sensor perception technology, WiFi signal-based human action recognition technology has the advantages of not infringing on privacy, not being affected by light, and low cost, and has become one of the most popular methods in the field of human action recognition. one. Compared with RSSI, CSI has finer granularity and sensitivity, and can perceive smaller changes in the channel, so it has better recognition accuracy for smaller movements such as breathing and gestures. [0003] There are still the following problems to be solved in cross-environment hu...

AI Technical Summary

Problems solved by technology

[0006] To sum up, although the HAR model using the transfer learning method or the small sample learning method still maintains a high accuracy rate in the new scene, the transfer learning requires a large number of action samples in the new scene to fine-tune the model. The increase in the number of types of actions leads to a significant decrease in the recognition rate

At the same time, after the small-sample learning and transfer learning methods learn to recognize new actions, the accuracy of the new model for the original action recognition is greatly reduced. It is only effective in the current scene and cannot achieve continuous learning.

The reason why these two types of methods forget the old task is that in the training of the new task, the connection weights between the neurons in the network must be adjusted for the new task, which will change the original structure adapted to the old task, resulting in performance drops rapidly on

Images