| Gesture recognition is an important research direction in the field of humancomputer interaction,which has many applications in smart cities,sign language translation and home automation As one of the current research hotspots,gesture recognition methods based on wireless signals have the characteristics of low cost,wide distribution and good scalability.In this thesis,single person gesture behavior is took as the research object and the gesture actions is analyzed through raw data processing,gesture feature extraction and network model construction to achieve feature fusion of continuous gesture actions and improve the recognition effect of the algorithm.Channel State Information(CSI),as a collection of wireless channel states,can be used to characterize the perceived changes in human motion during signal propagation,and is recorded in magnitude and phase.Considering its susceptibility to environmental and multipath propagation effects and the presence of noisy data,it is proposed to separate the key motion information from the noise-obscured information based on Principal Component Analysis(PCA)and Discrete Wavelet Transform(DWT)to construct gesture behavior features.Meanwhile,the existing segmental gesture feature extraction rules are difficult to describe the overall semantics of the movement process.Considering the continuity of background and action information when gestures change,it is proposed to integrate the features of the forward gesture action sequence with those of the reverse gesture action sequence using temporal weight sharing to achieve the analysis and recognition of the overall features of gesture behavior.In this thesis,it is proposed to use the combination of feature fusion and temporal modeling to build a BAGR(Bi-LSTM Attention Gesture Recognition)network to fuse forward and reverse gesture action sequences with bi-directional features to solve the limitations and shortcomings of analysis only based on the historical information of gesture actions.Firstly,principal component analysis is used to construct effective action features from the original data.Secondly,the gesture behaviors are extracted from shallow to deep temporal features through pyramidal Bi-LSTM network structure to achieve multi-level representation of gesture information.In addition,Channel Attention(CA)is introduced to assign weights to different feature extraction channels to enhance the model’s feature learning of key action information.Finally,Parametric Rectified Linear Unit(PRe LU)is introduced to update the activation function parameters adaptively to improve the robustness of the model.Meanwhile,a multi-model fusion-based approach is proposed to achieve deep learning of spatio-temporal features of gesture actions by building a CBGR(CNN BiGRU Gesture Recognition)network to solve the problem of modeling the mapping relationship of gesture features in spatio-temporal dimensions.Firstly,the gesture features of the low frequency part are extracted by discrete wavelet transform to realize the feature mining of the original data.Secondly,a multi-input network structure is constructed and the spatial features of gesture actions are extracted by progressive convolutional neural network,and it is combined with residual structure to strengthen the model fitting ability,and finally it is fused with the temporal features of gesture actions extracted by Bi-GRU to achieve the deep capture of spatio-temporal features of gesture behaviors.Gesture actions are trained and recognized by a temporal network model with data preprocessing,feature extraction,and fused attention mechanism.The experimental results show that the BAGR network and the CBGR network proposed in this thesis achieve 94.1% and 96.8% accuracy on the Widar 3.0 gesture action test set,respectively,which shows that the relevant algorithms achieve effective modeling of gesture behavior based on wireless signals and have strong adaptability. |
