Indoor Human Activity Recognition Based On Wi-Fi Channel State Information

Activity recognition,as an important medium in human-computer interaction,has been widely studied in intelligent home,medical care and other life fields.In the field of wireless perception,activity recognition technology based on Wi-Fi Channel State Information(CSI)makes up for the shortcomings of video surveillance technology that is easy to invade people’s privacy and wearable devices need to be worn in real time.In recent years,it has developed rapidly in the field of passive activity recognition.However,after changing scenes and adding new users,the activity recognition scheme based on Wi-Fi CSI has poor recognition accuracy due to the lack of sufficient activity samples and the change of activity data brought by the change of users and scenes.Therefore,this thesis proposes an indoor human activity recognition algorithm for new users and cross-scenes in Wi-Fi scenarios.The main contents are as follows:(1)Aiming at the problem that the accuracy of classification model decreases due to insufficient CSI activity data in new Wi-Fi scenarios and new users,this thesis proposes a method of CSI activity data amplification based on least square loss of cyclic adversarial network.First of all,this theis converts CSI signal data into signal graph,combines real activity data of original users and environment and ten percent real activity data of new users and new environment,and then amplifies CSI activity data of new users and new environment based on cyclic adversarial network.In order to stabilize the process of network learning,we use the least square loss to optimize the quality of CSI signal graph generated by generator.The experimental results show that in the human daily activity and gesture activity recognition tasks for new users and new environments,the proposed data amplification method combined with a variety of classification networks can improve the activity recognition accuracy by more than 15% compared with before amplification.(2)Aiming at the problem that the low accuracy of new user activity recognition in existing Wi-Fi scenarios,this thesis proposes a small classification network model based on dense convolution and boundary constraints for new user activity recognition.This thesis reconstructs dataset for the new user first,then designed a small benchmark networkbased on intensive convolution block and embedded in the boundary constraint module learning representative features of each type of activity.The features of the different users the same category activity space more gathered themselves together,and different activity categories between the characteristics of the spatial distribution is more spread out.The accuracy of the proposed new small user activity recognition algorithm is more than 3% higher than that of the advanced deep neural network on the public data set and the gesture activity and daily activity data set collected in thisthesis,and the number of parameters is only ten percent of the advanced classification network.(3)Aiming at the problem that the accuracy of activity recognition task based on Wi-Fi CSI decreases in cross-scene and cross-user situations,this thesis proposes a classification network model based on self-calibration crossover and hierarchical cascade attention mechanism.The self-calibration cross-attention module divides the feature map channel into self-calibration convolution operation and cross-attention weight learning to enhance the network learning of local feature information.The hierarchical self-attention module divides the feature map into equal regions,models the global feature relations in a hierarchical manner,and learns long-distance context information.The experimental results show that the proposed general activity recognition algorithm achieves more than 95% accuracy in the cross-scene and cross-user human daily activity and gesture activity recognition tasks.In conclusion,the proposed method of this thesis can achieve high recognition accuracy in indoor new user and cross-scene human activity recognition tasks based on Wi-Fi CSI.Among them,the small new user activity identification network can be applied to intelligent home and other human-computer interaction fields,and the general network identification model can provide a new solution for the future Wi-Fi CSI activity identification task.