Device-Free Indoor Localization Based on Kernel Dictionary Learning

In this article, we consider a kernel dictionary learning method to solve the device-free localization (DFL) problem. Generally, DFL is formulated as a sparse classification problem. However, localization performance is often unsatisfactory because different classes of localization data vectors are linearly indistinguishable and, thus, may be distributed in the same direction. Therefore, solving the DFL classification problem is not expected to provide a superb DFL performance using general dictionary learning. To improve localization accuracy, we propose kernel dictionary learning for DFL, which combines the kernel function with dictionary learning, to solve the problem of indistinguishable low-dimensional localization data and learn the dictionary applicable to a multivariate environment. The data are the first kernel transformed into a higher dimension, to make data more classifiable. Then, the pseudo-dictionary is obtained by kernel k-singular value decomposition (KKSVD). Finally, the pseudo-dictionary is used to localize the target by calculating the minimum residual. The experimental results show that KKSVD displays good noise immunity compared to KSVD and other sparse coding algorithms, such as orthogonal match pursuit and basis pursuit (BP), among others.