Learning Individual Models to Estimate the Walking Direction of Mobile Phone Users

Abstract

Pedestrian Dead Reckoning algorithms are commonly used to assist pedestrian navigation with handheld sensors. The estimation of the walking direction remains an important source of positioning error of mobile phone users since this direction may be different from the device’s pointing direction. A better understanding of human walking gait has enabled to produce new algorithms to mitigate the impact of the way the device is held in hand. WAISS algorithm is one of them. It is based on the study of horizontal hand accelerations and their modeling using Gaussian Mixture Models (GMM). However, ongoing search for universal modeling of handheld device carrying mode defeats the varying nature of human gait. This paper investigates the impact of individual gait characteristics and their modeling to improve the estimation of the walking direction. Different models are learned for curved and straight lines and varying GMM are proposed to account for inter-individual gait variations. This results in a reduced walking direction error with a 8.1° mean error to the 90th percentile computed for 3 subjects over a 1.5 km indoor/outdoor walk.