Complex-valued Deep Learning for WiFi-based Indoor Positioning: Method and performance

Inhalt:
Deep learning-based fingerprinting for indoor positioning based on channel state information (CSI) has yielded promising results. This paper introduces a complex-valued Convolutional Neural Network (CV-CNN) architecture explicitly designed for indoor positioning based on CSI. The proposed model is evaluated for implementation with the IEEE 802.11n and IEEE 802.11ax standards due to the the existing packages capable of extracting CSI readings in commercial off-the-shelf (COTS) devices. Our results indicate that the proposed CVCNN demonstrates enhanced capability for generalizing positions not included in the training reference points (RPs) compared to real-valued Convolutional Neural Network (RV-CNN). For the considered scenario, the proposed architecture achieves a positioning error of less than 1.25 meters with a probability of 90%. In comparison, RV-CNN architectures achieve an error of more than 3 meters in the same scenario.