A Transfer Learning Based CSI Indoor Localization Using GASF Image Construction

When the indoor environment changes, how to reduce the channel state information (CSI) measurement collection for off-line training has received much attentions. In this paper, a transfer learning based CSI indoor localization using Gram's corner field (GASF) image construction is proposed. The density-based clustering algorithm (DBSCAN) is proposed to remove abnormal CSI measurements and reduce noise. Then, the sequence of CSI amplitude information is transformed into the gray CSI image by GASF method. The linear mapping is used to form the final color CSI image which is used as fingerprint. This proposed algorithm collects a large number of CSI signals in the source domain, and the convolutional neural network (CNN) with attention mechanism is used to training the localization model in the source domain. In the target domain training, after the same data preprocessing of small number of CSI measurements, some model parameters are fine-tuned by the target domain CSI image. At last, the target domain localization model is obtained. Experimental results show that the proposed algorithm can not only achieve high-precision localization performance, but also reduce the training time.