Complex-Valued Convolutional Neural Network for Indoor Positioning

Deep learning has been recognized as an effective method for indoor positioning, particularly in dense non-line-of-sight (NLOS) environments. However, most existing neural networks designed for indoor positioning view the two constituent components of channel state information (CSI) as real-valued inputs, which may discard some useful information embedded in the original CSI. In this letter, we propose a complex-valued convolutional neural network for indoor positioning named CCPNet, which incorporates local and global feature extraction modules and fully exploits the correlation between real and imaginary parts through complex-valued convolution. The experimental results demonstrate that, compared to existing real-valued neural networks, CCPNet exhibits more than 20% improvement in positioning accuracy while maintaining a relatively small number of parameters.