Research On CSI Localization And Activity Recognition By Fusion Strategy Under Multi-Receivers Condition

With the widespread application of artificial intelligence technologies,the demand for target localization and activity recognition has increased dramatically.Due to its low cost and easy deployment,Wi-Fi based localization and activity recognition has received more and more attentions.In this paper,the application scenario is extended from single receiver to multiple receivers scenarios.the channel state information(CSI)based localization and activity recognition is studied by fusion strategy.The main contributions of this paper include:(1)Theoretical knowledge related to CSI based localization and activity recognition is studied.Firstly,the amplitude and phase information of CSI measurements are described.And then the convolutional neural networks(CNN)and distributed learning framework are introduced in detail.Finally,the hardware platform and software platform for CSI data collection are built.Through above discussions,they provide the solid foundation for the following research of the paper.(2)A localization and activity recognition algorithm by distributed learning framework and linear weighted decision fusion is proposed.In the off-line phase,the Hampel filter and Gauss filter are used to mitigate the noise effect of CSI amplitude measurements from multiple receivers.Then,the CSI fingerprint images are constructed from the time and frequency domains.For each receiver’s training data set,the position estimation and activity recognition models are trained by convolutional neural network(CNN)by a distributed learning framework.In the on-line phase,the pre-processed CSI fingerprint image is firstly sent to the position estimation and activity recognition models for the preliminary results of localization and activity recognition,respectively.Then,the linear weighted decision is proposed to fuse these preliminary results and obtain the final results.The proposed algorithm can accelerate the training time and convergence speed by using a distributed learning framework.The experimental results verify the efficiency of the proposed algorithm.(3)A localization and activity recognition algorithm by multi-task learning framework and adaptive entropy weighted decision fusion is proposed.In the off-line phase,after the data preprocessing,the CSI amplitude fingerprint images with corresponding position and activity labels are used to form the training datasets.For each receiver’s training data set,the localization and activity recognition tasks are jointly trained by a multi-task learning framework and obtain the position estimation and activity recognition models.In the on-line phase,the constructed CSI amplitude image is sent to different position estimation and activity recognition models to obtain preliminary estimation results.Finally,the different results are fused by adaptive entropy weighted decision fusion and obtain the final estimation results.The proposed algorithm can make use of the shared layer information through hard parameter based sharing and also reduce the risk of overfitting.The experimental results show that the proposed algorithm can obtain better localization and activity recognition results.