| With the development of artificial intelligence and wireless sensing technology,WiFi-based non-contact gesture sensing technology has been widely used in fields such as smart homes,game control,and rehabilitation therapy.It has gradually integrated into daily life,and the demand for human-computer interaction has also been continuously increasing,which has promoted the research and advancement of Wi-Fi gesture sensing.Although existing research methods can achieve high recognition accuracy for gesture motion sensing,there are problems with poor model perceptual universality and low recognition accuracy when there are insufficient training samples,changes in gesture motion direction,and changes in the scene and category of recognition.To address these issues,this thesis extracts the Channel State Information(CSI)data caused by gestures from Wi-Fi signal devices,and studies different feature extraction and classification recognition methods based on a small number of samples,different gesture motion directions,different scenes,and categories to achieve robust non-contact gesture motion sensing.This thesis focuses on the research of several commonly used gesture motions,and the specific work is as follows:(1)To address the problem of model overfitting due to a small number of annotated samples,and the decrease in recognition accuracy caused by changes in gesture motion direction,a gesture recognition method named CSI-VGR based on time convolutional neural network and geometric transformation was designed.Firstly,gesture actions were collected and denoised using median filters and Savitzky-Golay filters on Wi-Fi sensing devices.Then Principal Component Analysis algorithm was used for dimension reduction and principal component extraction,and Short-Time Fourier Transform was used to extract Doppler shifts and geometric transformation was used to expand the data samples.Afterwards,the encoded-decoded time convolutional neural network was designed to segment the data and the segmented activity data was input into the Temporal Convolutional Network for feature extraction.Finally,the Echo State Network ESN was introduced for gesture feature classification.Through experimental tests in two different real environments,the recognition accuracy of the CSI-VGR method reached 96.3%,achieving good recognition for a small number of gesture samples.(2)To address the problem of low perceptual accuracy and limited geometric transformation data augmentation caused by different gesture scenes and categories,a gesture recognition method named Wi-TCG based on transfer learning and generative adversarial network was proposed.Firstly,the collected CSI data was denoised and transformed into Doppler shift images.Secondly,the image data and corresponding category label data were used as input for conditional generative adversarial network to generate pseudo data and expand the sample data.Finally,the Res Net18 deep residual network trained from the source domain Widar3.0 dataset based on model transfer learning was fine-tuned with the network parameters in the new scene and new gesture categories in this article to recognize the collected gesture motion data.Wi-TCG was evaluated in two environments,office and meeting room,and the average recognition accuracy was 93.1%,achieving gesture recognition of different scenes and categories.(3)To address the limitations of sample data distribution in transfer learning and the difficulty in collecting and annotating a large number of training samples in deep learning,a gesture recognition method named Meta-WGR based on bidirectional gated recurrent unit and meta-learning was proposed.Meta-WGR learns classification experience from human actions in the Wi AR dataset and gesture actions in the Widar3.0 dataset through meta-learning and transfers the learned knowledge to our self-built gesture category dataset.Firstly,the dataset and collected CSI data were denoised and smoothed using a combined filter and transformed into Doppler shift.Then,the feature tensor of the image was extracted using a convolutional neural network with attention mechanism.Finally,the bidirectional gated recurrent neural network was used as the classification model,and the classification task of the two datasets was trained and learned in the meta-training phase.The obtained initialization parameters were used as input for meta-testing and finetuning and recognition in a new category task with a small number of different hand gesture samples.After multiple experiments,the recognition accuracy of Meta-WGR on hand gestures reached 92.5%,which can meet the interaction requirements of non-contact Wi-Fi sensing and achieve recognition of different categories and a small number of sample data. |
