Research On RFID-based Cross-scene Contactless Gesture Recognition

Radio Frequency Identification(RFID)has become a key technology in the field of contactless gesture recognition due to its light weight,ease of deployment and low interference from signals in the same frequency band.However,due to the characteristics of reflection,refraction and diffraction of wireless signals,the signals of the same action collected in different scenes may vary.This is because different scenarios,such as different users generating the action,different distances between the action and the sensing platform,or different deployment environments,can have different effects on the propagation of the wireless signal.Models trained in some scenarios often do not achieve satisfactory performance in new scenarios,and therefore need to improve the cross-scene capability of the models.The so-called cross-scene capability is to enable the model to still have high performance when the scenario changes.This thesis presents an in-depth study on how to improve the cross-scene capability of RFID-based contactless gesture recognition from two aspects: signal pre-processing and classification model construction,with the following research work.(1)To ensure that the perception model can obtain better performance,the received wireless signals need to be preprocessed first to obtain more accurate signal representations.In the preprocessing process,changes in background noise distribution caused by scene changes can lead to inaccurate segmentation of gesture signals;the limited time-frequency expression capability of common feature representations limits the learning capability of the model.To address the problem of inaccurate segmentation of gesture signals,this thesis designs a segmentation method based on fusion variance for signal segmentation.The variance of the active signal collected from different labels is subtracted from the noise variance of the scene,and the results are summed after exponential operation,and the signals of different labels are weighted and corrected according to the distribution of environmental noise.Equivalently,the threshold value of segmentation is dynamically adjusted according to the changing intensity of signal and noise,and the change of scene information is considered to assist in dividing a more complete signal.To address the problem that the capability of the existing feature representation needs to be improved,this thesis uses the simultaneous squeezed wavelet transform to obtain higher resolution time-frequency information.The simultaneous squeezed wavelet transform combines the ideas of wavelet transform and empirical modal decomposition to obtain instantaneous frequencies by calculating wavelet coefficients on the original signal,and then reorganizes its frequency information to obtain a finer time-frequency representation.Through experiments,it is proved that the preprocessing method proposed in this thesis can obtain clean gesture signals with high resolution very well.(2)In order to better enhance the cross-scene capability of the recognition model,the learning content of the model needs to be adjusted in a targeted manner.When the model learns,signal features with different information content require different learning capabilities,and unreasonable parameter settings will lead to waste of resources and overfitting;uncorrected extraction of information will lead to the model learning features containing a large amount of information that limits its cross-scene capability.To address the problem that signal features with different amounts of information require different learning capabilities,this thesis proposes a dual-channel feature extraction algorithm with different weights for phase and Received Signal Strength Indicator(RSSI)information learning by analyzing the backscattered signal characteristics of RFID tags,giving more learning weights to phases containing more information to better capture gesture signal features.To address the problem that the model learns a lot of scene-specific information and thus affects the cross-scene capability of the model,the continuous distance is discretized by considering the effect brought by distance variation,and a multi adversarial domain discriminator is designed to eliminate the information of different scenes.Through experiments,it is demonstrated that the recognition model proposed in this thesis can maximize the use of existing tag backscattering prior knowledge to improve the cross-scene capability of RFID-based contactless gesture recognition.To verify the effectiveness of the above method,a large amount of gesture data in multiple scenes is collected in this thesis,and a self-use dataset is constructed.The experimental results show that the algorithm proposed in this thesis improves the accuracy and robustness of RFIDbased cross-scene contactless gesture recognition,which provides more possibilities for the interaction methods between human and smart devices.