Deep learning-based mobile devices localization in GPS-denied environments

• Main Author: Yang, Shi Bo
• Other Authors: Wang Dan Wei
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Engineering
• Online Access: https://hdl.handle.net/10356/177257
• Institution: Nanyang Technological University

Location-based services (LBS) has revolutionized the way we navigate and manage assets by leveraging location data. Localization in LBS is an essential capability that empowers mobile devices with such functions. However, in environments where GPS signals are challenged, or denied, such as office corridors, hospitals, and supermarkets, traditional localization methods relying on perception sensors like camera and LiDAR often fall short due to repetitive and ambiguous settings. Meanwhile, some current research focuses on other localization methods such as the UWB-based approach, which is impeded by its numerous and expensive equipment requirements. This FYP project aims to estimate accurate positions of mobile devices in GPS-denied areas utilizing Wi-Fi access points (WAPs) and deep learning-based methods (i.e., LSTM). The main objective is to collect WAP data, design and train the LSTM model, and implement it for precise location estimation. First, the dataset was 70-30 split for training and testing, respectively, and then the training dataset was 60-40 split for training and validation, respectively. The designed LSTM network is composed of a linear rectification module used to linearly output the input features extracted from WAPs and an LSTM module with an attention layer. The LSTM module is designed for learning features and retaining long-term dependency information (i.e., position of the trajectory of the first n-1 moments). An attention layer is incorporated into the network to ensure it focuses on various features (i.e., different path types), enhancing generalization and averting gradient vanishing as well. Extensive experimental results demonstrated that the proposed system is capable of estimating the position of the mobile device with high accuracy and acceptable computational burden. Additionally, the proposed system achieves an accuracy of 1.2 meters, which is significantly more accurate than the conventional Wi-Fi localization methods with an error range of 3-10 meters. Meanwhile, the average training time of our model on a set of data (about 60 trajectories, 26160 WAPs) is 42 seconds, which is acceptable for real-world applications.