Lightweight convolutional neural network-based indoor localization of CSI images

Aiming at the problem of high computational complexity and large memory occupation of convolutional neural network (CNN), this paper proposes a lightweight CNN-based passive localisation method for channel state information (CSI) image fingerprints (LCNNLoc). In the offline training stage, the amplitude difference matrix and phase matrix are constructed into a three-channel feature image similar to "RGB"; at the same time, a lightweight CNN architecture is designed, the feature image is used as the input to train the framework, and the CNN model is saved as a fingerprint database at the end of training. In the online positioning stage, real-time position estimation was achieved using a probability weighted centroid method. The experimental results show that compared with the traditional method, LCNNLoc not only improves the positioning accuracy, but also reduces the algorithm running time consuming.