Effect of Adding Time Correlation to SVM-Based Motion Classification in Pedestrian Navigation

Abstract

In this article, we propose an approach to enhance zero-velocity-update (ZUPT)-aided inertial navigation systems (INSs) with a time series support vector machine (SVM) forecaster algorithm. The approach is based on the inclusion in ZUPT algorithm the time correlation of velocity threshold values based on classification of 19 distinct pedestrian activities determined from a foot-mounted inertial measurement unit. The classification enhances the traditional ZUPT-aided INS by first optimizing the threshold in the detector called stance hypothesis optimal detection and second adjusting zero-velocity measurement variances for each categorized locomotion type. Experimental validation involved three subjects, each conducting 10 trials of indoor navigation, encompassing activities, such as walking, fast walking, jogging, running, sprinting, walking backward, jogging backward, and sidestepping, over a nearly 100 [m] path. The trained time series SVM classifier achieved a 90.04% average classification accuracy, resulting in an improvement in navigation accuracy by a factor of 250 as compared to a standalone INS and by a factor of 3 as compared to a traditional ZUPT-aided INS solution. Comparable improvements in the vertical drift of the navigation solution have been also demonstrated.