Research On Human Motion Recognition Based On Channel State Information And Bi-LSTM

As a very fine-grained perception source,Channel State Information(CSI)of Wi -Fi signal can provide rich environmental change characteristics.It not only realizes more accurate behavior perception,but also has excellent non intrusiveness.Therefore,action recognition based on CSI plays an important role in the fields of family health,perva-sive scene perception and human-computer interaction.However,the current recognition methods add a large amount of redundant static information to the analysis,and often ig-nore the relationship between time context and state,which has some limitations for the recognition of actions that depend on long-time series.Meanwhile,the robustness is also low when facing scene changes,leading to repeatedly obtain and train a large number of target domain data.Therefore,in view of the above problems,the main research contents of this paper mainly include the following three aspects:1.Action extraction method based on mean absolute deviation:By studying the mod-eling method of channel state information,this paper extracts the base signals such as amplitude and phase difference in the frequency domain of channel subcarriers as iden-tification features,extracts the start and end points of activities from the time sequence based on the standard deviation of mean absolute deviation of amplitude,and removes the forward and backward redundant static information in the sample,consequently reduce the feature extraction dimension in time sequence during the recognition process.2.Bi-directional temporal feature recognition method:Bi-LSTM model based on deep learning is used to extract features Bi-directionally.Time-domain down sampling,Fast-LOF outlier detection and wavelet threshold denoising are carried out successively to re-duce and integrate the feature dimension.The amplitude and phase features are combined,and the attention mechanism is applied to assign different weights to the features according to the recognition contribution,so as to improve the effect of the model.3.Learning method based on feature transfer:On the basis of two-way temporal fea-tures,this paper studies the transfer learning method to solve the classification problem of actions in unknown fields by using the experience of trained fields.On the basis of freezing the appropriate CNN feature extraction layer,the multi-core maximum mean dif-ference method is used to optimize the full connection layer,minimize the difference of the two domains,increasing adaptability of the nomadic model within the target domain,and reduce the large time consumption of retraining the model when the scene changes.In this paper,the comparative experiments of different recognition methods are car-ried out on YSFHA-7 and Wi-AR public datasets.The experimental results show that the recognition method of two-way timing features with attention mechanism improves the recognition effect of timing dependent actions.And when the scene changes,the learning results of the source can be transferred to the target,and it is found that using a smaller learning rate has a better learning effect,which also provides a new idea for the general-ization performance of CSI recognition domain.