CoMuTe: A Convolutional Neural Network Based Device Free Multiple Target Localization Using CSI
With the growth of Internet-of-Things (IoT), Location based Services (LBS) are gaining significant attention over the past years. Location information is one of the important ingredients for many LBS where the system requires to localize multiple targets in indoor setting. As an emerging technique, device-free localization (DFL) is promising to localize the target without attaching any transceivers. In this paper, we propose CoMuTe, the first convolutional neural network (CNN) based device free multiple target localization leveraging the Channel State Information (CSI) from multiple wireless links. The system represents the CSIs as Multi-link Time-Frequency (MLTF) image by organizing them as time-frequency matrices and utilize these MLTF images as the input feature for CNN network. The CoMuTe models multi target localization as a multi label spot classification approach under the assumption that each MLTF image is associated with multiple labels/spots. The localization is performed with a training stage and a localization stage. In the training stage, the CSI based MLTF images are constructed with single target at each location. The constructed images are used to train the CNN via a gradient based optimization algorithm. In the localization stage, the test MLTF image obtained for targets at multiple spots is fed to the CNN network and locations of the targets are calculated using sigmoid activation function in the output layer under the multi label classification framework. Extensive experiments are conducted to select appropriate parameters for the CNN architecture as well as for the system design. The experimental results demonstrate the superior performance of CoMuTe over existing multi target localization approaches.