RoTrax: A Transformer-Based Framework for High-Accuracy IMU Trajectory Estimation in GPS-Denied Environments

Achieving the reliable positioning in various transportation modes is difficult when the global positioning system (GPS) or alternative satellite-based navigation signals are unavailable. In such cases, low-cost inertial measurement units (IMUs) provide the continuous sensing of motion; however, due to velocity instability and the accumulation of noise, drift gets added over time, resulting in poor positioning. In this study, a novel Transformer-based deep learning framework called robust trajectory reconstruction using Transformer (RoTrax) is introduced for enhancing the reliability of IMU-based navigation. This framework adjusts sensor bias correction, signal denoising, and coordinate alignment using a dedicated preprocessing pipeline. Compared with a recurrent architecture-dependent deep learning approach, RoTrax takes advantage of the Transformer attention mechanism to increase the temporal dependencies for reducing the accumulated drift effects. To evaluate the RoTrax performance, six transportation modes (walking, cycling, two-wheeler, three-wheeler, car, and bus), state-of-the-art comparison, and cross-device validation were considered. The results indicate that RoTrax is outperforming LSTM-based baseline models and achieving comparable performance with state-of-the-art Transformer-based IMU navigation methods. The cross-device validation further demonstrates the robustness and generalization capability of the proposed framework. This work provides the validation for RoTrax as a potential candidate for reliable positioning through a broad range of user mobility and conditions. © 2026 IEEE.