WiFi Fingerprinting Indoor Localization Based on Graph Attention Network for Multi-Floor Environments

Abstract

Indoor localization is a key technology for implementing location-based services. There are various location-based service applications such as healthcare, navigation, and tracking. We can conduct outdoor localization easily with the help of the Global Navigation Satelite System, which utilizes satellites in space. However, Global Navigation Satelite System does not work for people indoors since the signal from the satellites is blocked by the building's concrete wall. Therefore many technologies other than satellites have been considered for implementing the indoor localization system. Many researchers have focused on WiFi since the wireless access points that provide the wireless local area network are installed in most buildings. The wireless access points can act as satellites in the Global Navigation Satelite System. In other words, wireless access points can be the infrastructure for indoor localization. In addition, most people carry a mobile device like a smartphone equipped with WiFi network interface cards. Thus we do not have to buy or install additional hardware, so the WiFi-based indoor localization systems are universal. WiFi fingerprinting has been one of the most practical approaches for implementing an indoor positioning system because received signal strength can be measured in off-the-shelf devices. WiFi fingerprinting-based indoor localization methods use the received signal strength from all wireless access points as features for predicting the location labels. Many machine learning algorithms and neural networks are used to make prediction models. Existing methods design the indoor localization model using various machine learning methods and neural network architectures, including the k-nearest neighbor algorithm, support vector machine, recurrent neural network, and convolutional neural network. Among them, convolutional neural network-based models show outstanding performance. However, the fingerprint data has a weak locality because the received signal strength from one wireless access point and the received signal strength from the other are weakly correlated. Thus CNN may not be the best neural network architecture for WiFi RSS fingerprinting-based methods. In this dissertation, We propose GraphLoc, the WiFi fingerprinting-based indoor localization method that uses a graph attention network to design the indoor localization model. We represent the fingerprint dataset as a graph structure whose nodes represent the reference points. The nodes whose fingerprint data are similar are connected via edge. If the fingerprint of one node and its neighbors are similar, the location labels of one node and its neighbors are likely to be similar. Therefore the convolutional operation with the graph representation of fingerprint data can extract strong spatial local features. We propose the method to represent the fingerprint data as a graph structure and design the multi-floor indoor localization model using graph attention networks. To evaluate the performance of GraphLoc, we performed the performance evaluation with the WiFi fingerprint datasets. The evaluation results show that GraphLoc outperforms the benchmark methods that used regular fingerprint data other than a graph.