WiT: Wireless Tracking With Dual-Modality Transformers

作者：Xiaopeng Zhao, Guosheng Wang, Lei Yang · 发表于：IEEE Transactions on Mobile Computing · 年份：2025 · DOI：10.1109/tmc.2025.3636924 · 被引用次数：1 · 研究领域：Indoor and Outdoor Localization Technologies、Millimeter-Wave Propagation and Modeling、Robotics and Sensor-Based Localization

Deep learning-based wireless indoor localization has emerged as a promising solution to real-world deployment challenges. However, existing research still relies heavily on high-quality training samples and is susceptible to environmental dynamics. To address these issues, we introduce WiT, a dual-modality model that integrates Radio Frequency (RF) signals and Inertial Measurement Unit (IMU) readings to robustly track the location of mobile devices. WiT enhances accuracy through cross-modality validation, mitigating the impact of environmental fluctuations. We implement a history-enabled tracking method to resolve synchronization issues among different modalities. Additionally, we propose a semi-supervised training approach to reduce the demand for ground truth labels. We train and validate WiT on a large-scale dataset comprising approximately 1.7 million samples across 54 scenarios. Our results demonstrate that WiT outperforms state-of-the-art solutions, achieving improvements of at least 30% on the Radio Frequency Identification (RFID), Wi-Fi, and Bluetooth Low Energy (BLE) modalities. Furthermore, integrating both IMU and BLE modalities yields nearly a 50% improvement in accuracy compared to single-modality input.