A Transfer Learning Framework Based CSI Localization Using MTF Image Construction and KL Divergence Method

Indoor localization technology has been received much attentions. Channel state information (CSI), a parameter that can be extracted from Wi-Fi signals, has been widely studied in deep learning-based indoor localization system. However, due to the complexity of indoor environments and the high cost of data collection, the learning performance is not insufficient. In this paper, a transfer learning framework is proposed for CSI localization. First, the CSI measurement sequences in the source domain is proposed to expand the training data in the target domain. Using the Kullback-Leibler (KL) divergence value as a weighting parameter, the CSI amplitude sequences in the source and target domains are added linearly to obtain the expanded CSI training data in the target domain. Then, the Markov Transition Field (MTF) algorithm is proposed for CSI image construction, Next, a two-step model-based transfer learning is proposed. After the parameters initialization of a pre-trained model in the source domain, the parameters of the localization model are fine-tuned in the first step by the data in the target domain. Then, for the second training step, and parameters of the fully connected layer are fine-tuned by freezing the convolutional and pooling layers. Experimental results verify the efficiency of the proposed algorithm.