| In recent years,with the rapid development of wireless sensing technology and the rapid emergence of smart devices,people’s lives have become more and more convenient.Excellent human-computer interaction ability is a very important requirement in artificial intelligence.The demands for human-computer connection are increasing,especially in the digital age,as people enter the rapid growth of artificial intelligence technologies.Among wireless signals such as infrared detection technology,Wi-Fi technology and Zig Bee,the wireless sensing technology based on Wi-Fi signal is widely deployed because of its wide device deployment,the security of protecting user privacy,the ubiquity of no light,and the low deployment price.It has very important research significance in the perception and recognition based on wireless signals,so it also has great potential in human-computer interaction.Under this background,more and more researchers have begun to study the wireless sensing technology based on Wi-Fi signal.Researchers can extract the state information(Channel State Information,CSI)in the Wi-Fi channel for indoor positioning,intrusion detection,people counting,gesture recognition,breathing detection,geriatric care and behavior perception and many other fields.However,due to the high correlation between CSI data and environmental information,the traditional machine learning method modeling model is crude and cannot achieve refined modeling of human gesture activities,resulting in unsatisfactory trained models.At the same time,the current Wi-Fi-based gesture recognition Most of the methods are based on scene-specific training,which means that gesture recognition accuracy drops dramatically when new scenes appear.Research on cross-domain gesture recognition technology has theoretical significance and practical value for multi-scene under the intelligent human-computer interaction.This paper deeply studies the Wi-Fi gesture recognition technology,further improves the accuracy of cross-domain gesture recognition,and conducts a large number of related experiments,and proposes two cross-domain gesture recognition methods,one based on manual extraction of domain-independent features and one based on automatic A method adapted to focus on extracting domain-independent features.The main work contents are as follows:Methods based on feature extraction by hand.The body coordinate velocity spectrum(BVP),which is independent of the domain environment,is extracted from the Wi-Fi signal based on its characteristics,and a residual 3D network 3D RAN network based on attention mechanism is designed for the domain-independent feature BVP to extract the BVP’s space-time.Features,realize cross-domain gesture recognition,and further improve the accuracy of gesture recognition.The model is tested on a public dataset,and the findings show that in the case of 6 different gestures in 3 situations,the average cross-domain identification accuracy based on BVP reaches87.4 percent.2.A method for extracting domain-independent features based on adaptive focusing.Aiming at the cumbersomeness of manually extracting features,a deep residual network DRAN is proposed to extract domain-independent features by adaptive focusing.It can achieve the best cross-domain recognition performance,and an effective CSI visualization method is proposed.The CSI stream is fused into a time-series image,providing a more fine-grained gesture description,while simulating the human attention mechanism,effectively connecting the attention mechanisms of multiple dimensions,and cooperating with the basic residual network Res Net to dynamically focus on the spatial and temporal dimensions.Gesture information cues,the model is evaluated on a public dataset,and the experimental results show that the cross-scene gesture recognition accuracy based on the deep residual network reaches 92.6% in the case of 6 different gestures in 3 scenes.And compared with cross-scene gesture recognition methods in recent years,our method has a higher recognition rate. |
PDF Full Text Request