Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learned Inertial Odometry for Cycling Based on Mixture of Experts Algorithm

Hao Qiao, Yan Wang, Shuo Yang, Xiaoyao Yu, Jian kuang, Xiaoji Niu

TL;DR

This work tackles the problem of robust bicycle localization under GNSS limitations by introducing a lightweight learned inertial odometry approach based on a Mixture-of-Experts (MoE) architecture. The MoE framework combines a gating network with multiple experts (built on a ResMLP backbone) to estimate 3D velocity from IMU measurements and orientation, while a tightly coupled error-state EKF fuses this velocity with raw IMU data to recover pose, velocity, and biases. Training optimizes velocity accuracy with MSE and NLL losses while enforcing balanced expert utilization through an auxiliary loss and an expert-capacity constraint. Experimental results on a 25+ hour, multi-participant bicycle dataset show the MoE method achieves comparable localization accuracy to LLIO but with substantial reductions in parameters (~64.7%) and computation (~81.8%), enabling feasible deployment on mobile devices and better robustness on unpaved surfaces. The work highlights a practical path toward efficient, data-driven inertial odometry for cycling and points to future directions such as knowledge distillation and mobile evaluation for broader applicability.

Abstract

With the rapid growth of bike sharing and the increasing diversity of cycling applications, accurate bicycle localization has become essential. traditional GNSS-based methods suffer from multipath effects, while existing inertial navigation approaches rely on precise modeling and show limited robustness. Tight Learned Inertial Odometry (TLIO) achieves low position drift by combining raw IMU data with predicted displacements by neural networks, but its high computational cost restricts deployment on mobile devices. To overcome this, we extend TLIO to bicycle localization and introduce an improved Mixture-of Experts (MoE) model that reduces both training and inference costs. Experiments show that, compared to the state-of-the-art LLIO framework, our method achieves comparable accuracy while reducing parameters by 64.7% and computational cost by 81.8%.

Learned Inertial Odometry for Cycling Based on Mixture of Experts Algorithm

TL;DR

This work tackles the problem of robust bicycle localization under GNSS limitations by introducing a lightweight learned inertial odometry approach based on a Mixture-of-Experts (MoE) architecture. The MoE framework combines a gating network with multiple experts (built on a ResMLP backbone) to estimate 3D velocity from IMU measurements and orientation, while a tightly coupled error-state EKF fuses this velocity with raw IMU data to recover pose, velocity, and biases. Training optimizes velocity accuracy with MSE and NLL losses while enforcing balanced expert utilization through an auxiliary loss and an expert-capacity constraint. Experimental results on a 25+ hour, multi-participant bicycle dataset show the MoE method achieves comparable localization accuracy to LLIO but with substantial reductions in parameters (~64.7%) and computation (~81.8%), enabling feasible deployment on mobile devices and better robustness on unpaved surfaces. The work highlights a practical path toward efficient, data-driven inertial odometry for cycling and points to future directions such as knowledge distillation and mobile evaluation for broader applicability.

Abstract

With the rapid growth of bike sharing and the increasing diversity of cycling applications, accurate bicycle localization has become essential. traditional GNSS-based methods suffer from multipath effects, while existing inertial navigation approaches rely on precise modeling and show limited robustness. Tight Learned Inertial Odometry (TLIO) achieves low position drift by combining raw IMU data with predicted displacements by neural networks, but its high computational cost restricts deployment on mobile devices. To overcome this, we extend TLIO to bicycle localization and introduce an improved Mixture-of Experts (MoE) model that reduces both training and inference costs. Experiments show that, compared to the state-of-the-art LLIO framework, our method achieves comparable accuracy while reducing parameters by 64.7% and computational cost by 81.8%.
Paper Structure (15 sections, 22 equations, 2 figures, 1 table)

This paper contains 15 sections, 22 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. System Overview
  3. Algorithm Description
  4. MoE model
  5. Network architecture
  6. Extended Kalman filter
  7. System State Definition
  8. State Propagation
  9. Measurement Update
  10. experiments
  11. Set Up
  12. Dataset
  13. Evaluation Metrics
  14. System Performance
  15. Conclusion

Figures (2)

  • Figure 1: The data acquisition device mounted on a bicycle handlebar.
  • Figure 2: Trajectory visualization under four different conditions.