Title: FDIO: Frequency Decomposed Inertial Odometry

Abstract:

Pedestrian inertial odometry (PIO) offers a promising solution for consumer-grade localization by estimating a person’s autonomous movement using solely acceleration and angular velocity data from an inertial measurement unit (IMU). However, in dual-device acquisition scenarios, IMU data captured by a handheld mobile device is inherently composite. The global motion of the human torso is intertwined with perturbations caused by local limb movements, complicating the accurate modeling of human motion.

To overcome this challenge, this study introduces Frequency Decomposed Inertial Odometry (FDIO). The proposed approach begins by separating input IMU signals into high-frequency and low-frequency components via a Laplacian pyramid. It then employs a Mamba module to capture long-range motion patterns within the low-frequency data, while a multi-scale convolution module is utilized to extract fine-grained local dynamic features from the high-frequency data.

Evaluations across five public PIO datasets demonstrate that FDIO attains an average absolute trajectory error of 3.221 meters and an average relative trajectory error of 2.550 meters. These figures represent reductions of 33.3% and 16.7%, respectively, when compared to the RoNIN ResNet baseline, thereby confirming the efficacy of the frequency decomposition strategy. To the best of our knowledge, this research represents one of the initial attempts to integrate Mamba architectures and frequency decomposition techniques into the field of inertial odometry.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC