Zero Velocity Detection for Pedestrian Inertial Navigation Based on Spatiotemporal Feature Fusion

Abstract

Zero velocity detection is a critical component in zero velocity update (ZUPT)-aided foot-mounted pedestrian navigation systems. Robust and accurate zero velocity detection significantly enhances the precision of pedestrian trajectory estimation. Existing zero velocity detectors based on fixed threshold and gait cycle segmentation techniques struggle to adapt to the complexity and variability of human motion. To address this issue, we propose an adaptive zero velocity detector based on deep learning. The raw inertial data possess spatial features with significant differences and temporal features that conform to certain patterns. This detector utilizes a contrastive learning (CL) network and a long short-term memory (LSTM) neural network (NN) to extract the spatial and temporal features of the inertial data, respectively. Experimental results demonstrate that the detector can achieve adaptive zero velocity detection and improve trajectory estimation accuracy regardless of individual differences or motion types. In two indoor experiments, the 2-D position error is 0.410 m for a mixed walking and running path, and the 3-D position error is 0.546 m for a mixed walking, running, and up/down stairs path.