This is the official repository for the paper: "iMoT: Inertial Motion Transformer for Indoor Navigation", accepted at AAAI 2025. This paper introduces an innovative inertial transformer architecture designed to model human motion uncertainties across velocity segments in a multimodal learning manner, thereby significantly enhancing trajectory reconstruction. Typically, during encoding, we first present Progressive Decoupler Series (PDS) to break down complex IMU signals into more interpretable components, facilitating the capture of specific motion patterns. Following this, we design Adaptive Positional Encoding (APE), a flexible and relaxed positional encoding mechanism that effectively learns discrepancies between tokens based on modality content. In addition, inspired by the Particle Filter, we introduce a set of learnable query motion particles corresponding to specific motion modes, which are progressively adjusted during the decoding steps to explain uncertainties in motion.
The code supporting the findings of the paper "iMoT: Inertial Motion Transformer for Indoor Navigation" will be made publicly available soon. Stay tuned for updates in this repository!
Coming soon! The official paper link will be provided here as soon as it becomes available.
For questions or collaborations, please feel free to reach out to us via email: [email protected] or [email protected]