Domain Adversarial Graph Convolutional Network Based on RSSI and Crowdsensing for Indoor Localization

In recent years, due to the wider WiFi coverage and the popularization of mobile communication devices, the technology of indoor positioning using WiFi fingerprints has been rapidly developed. Currently, most supervised methods need to collect a large amount of data to construct fingerprint datasets, which is labor-intensive and time-consuming. In addition, many studies focused on the ideal laboratory environment and lack the consideration in the practical application environment, especially in the scenario of multiple large multi-floor buildings. To solve these problems, we proposed a novel WiDAGCN model which can be trained by a few labeled site survey data and unlabeled crowdsensing WiFi fingerprints. To comprehensively represent the topology structure of the data, we constructed heterogeneous graphs according to the received signal strength indicators (RSSIs) between the waypoints and WiFi access points (APs). Moreover, previous WiFi indoor localization studies rarely involved complete graph feature representation, thus we use graph convolutional network (GCN) to extract graph-level embeddings. There are also some difficult problems, for example, a large amount of unlabeled data that cannot be applied to a supervised model, and the existence of multiple data domains leads to inconsistency in data distribution. Therefore, a semi-supervised domain adversarial training scheme was used to make full use of unlabeled data and align the data distribution of different domains. A public indoor localization dataset containing different buildings was used to evaluate the performance of the model. The experimental results show that our system can achieve a competitive localization accuracy in large buildings such as shopping malls.