Abstract
(type = abstract)
Learning important human contextual information is one of the fundamental features to establish a smart home environment. Location information is one of the important ingredients for many context-aware applications where the system requires to localize targets in an indoor setting. Most existing work, however, acquires contextual information in an obtrusive manner – they may require subjects to carry mobile devices, or rely on self or peer report to report data. As an emerging technique, device-free localization (DFL) is promising to localize the target without attaching any transceivers. Specifically, wireless sensing based device free localization is getting increased attention due to its ability to potentially support a broad array of applications including elder care, well-being management, and latchkey child safety, etc. In this work, we investigate the device free, unobtrusive approach for indoor localization using channel state information (CSI) leveraging machine learning techniques. In particular, our work mainly focuses on developing a multiple target localization system in a device free setting by formulating the localization problem as a machine learning based location/spot classification problem in order to achieve a robust, low cost, yet highly accurate localization system.
Multiple machine learning techniques have been explored in this work that can help us to learn human location at indoors by analyzing the location specific feature pattern of channel state information. First,
we designed MaLDIP, a support vector machine (SVM) algorithm framework to localize a single target through supervised classification problem. The system works by utilizing frequency diversity and spatial diversity properties of CSI at target location by correlating the impact of human presence to certain changes on the received signal features. However, accurate modeling of the effect of a subject on fine grained CSI is challenging due to the presence of multipaths. We propose a novel subcarrier selection method to remove the multipath affected subcarriers to improve the performance of localization. Finally, we select the most location-dependent features from channel response based upon the wireless propagation model for SVM based classification approach. The proposed approach results in much higher accuracy compared to the state-of-the-art localization approaches.
Second, we present a discriminant learning approach for two different indoor localization systems using CSI amplitude and phases difference between receiver antennas, respectively. We investigate a Canonical Correlation Analysis based feature fusion technique using only CSI amplitude features in order to incorporate discriminative features of CSIs for localization. We further improved the performance by exploiting multi-view learning approach in the training stage to utilize diversity from different AP's (view) CSI data. Moreover, in order to exploit the complete wireless propagation features, we utilize bi-modal features of CSI in terms of both amplitude and phase, where we propose a relatively stable phase feature by considering phase difference between receiver antennas. To design the multi-view learning algorithm, we implement GI$^{2}$DCA, a Generalized Inter-view and Intra-view Discriminant Correlation Analysis. GI$^{2}$DCA is a discriminative feature extraction approach that incorporates inter-view and intra-view class associations, while maximizing pairwise correlations across multi-view data sets. Finally, we use an Eucledean Distance based similarity measure to find the best match to localize a subject. Leveraging CSI data in terms of both amplitude and phase through multi-view discriminant learning approach, the proposed system can estimate location with high accuracy, which outperforms other benchmark approaches.
One of the great challenges for real-world application of indoor localization is to localize multiple targets and multiple targets scenario is usual in practical application. In this work, we present CoMuTe, a convolutional neural network (CNN) based device free multiple target localization leveraging the frequency diversity of CSI. Being a supervised method, only one group of weights for all the training locations are required in this approach, which is different from our prior works that requires training features for each training location. Specially, we represent the CSI amplitude dynamic as multi-link time frequency (MLTF) image by organizing them as time-frequency matrices for multiple wireless links and utilize these MLTF images as the input feature for CNN network. We model multi target localization as a multi label spot classification approach, where each MLTF image is assumed to be associated with multiple labels/spots. By exploiting convolutional layers with fully connected layers, the system can automatically extract the discriminant features of the MLTF images, to obtain training weights. We perform localization with a training stage and a localization stage. The CSI based channel image is collected for a number of training spots and used to train the CNN via adaptive learning rate optimization algorithm in the training stage. In the localization stage, the MLTF image collected for targets located at multiple spots is fed to the CNN network and the localization result is calculated using sigmoid activation function in the output layer under the multi label classification framework. The proposed system can achieve centimeter level location accuracy, which outperforms other benchmark multi target localization approaches.