Research On High Robust Motion Recognition Technology Based On Channel State Information

With the rapid development of technology,human-computer interaction is no longer a dream that only exists in science fiction movies.Motion recognition based on Wi Fi technology is an important part of human-computer interaction.Wi Fi wireless signals are divided into received signal strength indicator and channel state information.Among them,more fine-grained information can be identified based on the channel state information.Therefore,this paper will conduct research on high-robustness motion recognition technology for channel state information.The motion recognition process includes two phases which are offline and online.The input data is used in the offline phase for model training while online phase is responsible for final recognition.However,most studies do not take into account the fact that users trained in the offline phase and users identified in the online phase may not be the same person.Since different users will not completely agree even if they do the same action,and this inconsistency is more obvious in the channel state information that is easily affected by the environment,the difference between users directly affects the accuracy of the recognition.Aiming at the action recognition of non-identical users in offline and online stages,a method combining convolutional neural network and SVM classifier is proposed to jointly build a recognition training model.First,the collected CSI data was filtered out of outliers using the LOF algorithm,then sent to a convolutional neural network for feature extraction,and finally the SVM classifier was used for motion recognition.On the other hand,for the same user,even if the same action is performed in different situations,the actions will not be consistent due to issues such as amplitude and direction,which will also cause inconsistencies in channel state information and reduce the accuracy of recognition.Aiming at the same user's action recognition in offline and online phases,a method of matching stable features is proposed.This method is mainly based on human habits.Since every action performed by the user will add its own habitual characteristics,some of the same actions performed at different times are unchanged or little changed.The key point to solve this problem is to find stable characteristics.Secondly,because different subcarriers are affected differently by the channel,in order to improve the overall performance,the information of the key subcarriers that are greatly affected by the channel is extracted for identification.Then use PCA to further reduce the dimension and denoise the key subcarriers,and finally use random forest to classify the action.Experimental results show that extracting stable features can indeed improve the accuracy of recognition.