MSIN: An Efficient Multi-head Self-attention Framework for Inertial Navigation

Inertial Measurement Unit (IMU) makes an outstanding contribution to indoor inertial navigation in the era of ubiquitous computing, as it is widely integrated into portable devices. Many prominent works have been proposed by taking gyroscope and accelerometer readings as input to estimate the velocity and orientation. However, most of them focus on the local features of IMU (i.e., single sensor temporal feature or local spatial feature), eventually leading to drift on the trajectory. In this paper, we propose a robust model to mitigate the problem of jitters and drifts in trajectory prediction by exploiting the spatial dependence in accelerometer and gyroscope readings, as well as the contextual relation in motion terms through in-depth analyses of IMU readings. In particular, we design a framework (MSIN) to fuse the local spatial dependence of multiple sensors and incorporate the local spatial and global temporal features by using the multi-head self-attention mechanism. We have conducted extensive experiments on two public datasets and the results show that MSIN achieves a significant improvement in RTE (Relative Trajectory Error) performance, with improvements of up to 6.14% and 15.19% over state-of-the-art methods for RoNIN-Unseen and RIDI-Unseen, respectively.