DeepMetricFi: Improving Wi-Fi Fingerprinting Localization by Deep Metric Learning

The Wi-Fi RSSI fingerprinting method is one of the mainstream indoor localization solutions for its reliable positioning accuracy and ubiquitous infrastructure. The basic assumption is the location distance of the indoor environment can be estimated by signal distance based on the radio propagation models. However, the estimation could fail for the influence of the indoor environment, such as the multipath effect. Although the recent methods utilize machine learning techniques to improve the representation of the signal distance, most of them ignore the spatial information of the indoor environment where fingerprints are collected. In this article, we propose a deep metric learning-based Wi-Fi RSSI fingerprinting localization method aiming to learn effective RSSI features under the constraints of the reference point (RP) local structure to ensure the consistency of the location and signal distances in the indoor environment. First, we compute the path distance between the RPs to construct the positive and negative pairs from the fingerprints as the input. Then we design the deep metric learning model and minimize the triple loss by stochastic gradient descent (SGD). Finally, we extract the features of the online RSSI and conduct the localization with the features of the radio map by the WKNN method. In the experiment, the method is evaluated in a real scene with various regions, which may bring the challenge for localization. The results prove our method achieves better performance compared with the state-of-the-art.