Pedestrian Inertial Navigation Using Locomotion-Type-Specific Neural Networks

Abstract

Accurate pedestrian navigation in Global Navigation Satellite System (GNSS)-denied environments remains challenging. Pedestrian positioning systems should be self-contained and constrained by size, weight, and power. In some applications, such as defence, these systems should additionally be passive and difficult to detect. Inertial Measurement Unit (IMU)-only navigation is attractive in this setting, but direct integration of inertial measurements leads to rapid drift due to sensor noise and bias. Recent data-driven IMU-only pedestrian inertial-odometry approaches reduce this drift by using neural networks to provide short-term motion pseudo-measurements that are fused with inertial propagation in a filtering framework. However, human locomotion is heterogeneous, and a single model trained on all locomotion types may not perform equally well across different motion types.

This thesis investigates how estimated locomotion-type information can be incorporated into data-driven IMU-only pedestrian inertial odometry to improve state-estimation accuracy and uncertainty consistency. The proposed motion-aware learned-filter method combines a Convolutional Neural Network (CNN)-based locomotion-recognition module, motion-specific velocity-regression networks, and an Extended Kalman Filter (EKF)-based fusion architecture. The locomotion-recognition module classifies the current locomotion type from short IMU windows. The network corresponding to that class then predicts a body-frame velocity pseudo-measurement and its associated uncertainty for the EKF update. Two types of method variants are also investigated. The first uses the full probability distribution over locomotion classes in the EKF update. The second adapts the training settings separately for each locomotion class.

The results show that locomotion type can be recognized reliably from short IMU windows, although the classifier confidence decreases during transition periods. The motion-specific velocity-regression models improve neural-network prediction accuracy for selected locomotion types, especially running, sideway walking, and stairs up and down. The clearest improvement after EKF-based fusion is observed for running, where the average trajectory root mean square error (RMSE) over the tested trajectories decreases from 10.328 m to 8.063 m, corresponding to an improvement of approximately 22%. In some cases, such as sideway walking, selecting the motion-specific model using the ground-truth locomotion label does not necessarily produce the best trajectory estimate.

The investigated motion-specific training adaptations further reduce prediction errors for some motion-specific velocity-regression models, although the magnitude of the improvement depends on the locomotion type. The method variants that use the full probability distribution over locomotion classes have only a marginal effect when a single isolated transition occurs. However, on the dataset with more frequent locomotion changes, combining predictions from all motion-specific models reduces the average trajectory RMSE by approximately 6% compared with the proposed motion-aware learned-filter method.