Research On Key Techniques Of Dynamic Hand Gesture Recognition Based On 1D-CNN In WiFi Environment

With the continuous development of Internet of Things(IoT)technology and artificial intelligence,dynamic hand gesture recognition technology based on WiFi signal is expected to provide a new human-computer interaction(HCI)method.In recent years,although the hand gesture recognition system using Channel State Information(CSI)has made great progress,it cannot be directly extracted from most commercial network cards.The easily acquired Received Signal Strength Indication(RSSI)can only recognize simple hand gestures such as waving in the current research.In order to increase the upper limit of the system's ability to recognize hand gestures,this paper utilized data from multiple independent WiFi nodes to enable the system to recognize seven complex dynamic hand gestures.In order to realize the hand gesture automatic detection,this paper analyzed the principle of hand gestures affecting the channel,proposed a complete automatic hand gesture detection algorithm.The false trigger detection algorithm can effectly elimiate the false triggers and the hand gesture detection accuracy can reach 91.38%.The system recognize the start and end points of hand gestures by a state machine,combined with linear scaling algorithm the system can adapt to different hand gesture speeds of the user.Finally,in order to further improve the recogniton accuarcy,this paper analyzed the errors of the start and end points of the hand gesture segmentation obtained by the detection module and proposed a recognition architecture based on One-Dimensional Convolutional Neural Network(1D-CNN)and two data collection strategies:hand gesture extending and hand gesture shifting.The 1D-CNN architecture can effectively utilize the principle of shifting invariant detection of convolutional neural networks,eliminating the effects of automatic detection errors,and the two data acquisition strategies can collect more abundant samples to further improve the system performance.Through the experimental evaluation,the average recognition accuracy of the system is 93.03%,which is superior to other traditional machine learning algorithms.