Error Analysis of PDR System Using Dual Foot-mounted IMU

Abstract

In this paper, we analyze the position errors of the pedestrian dead reckoning (PDR) system using foot-mounted IMU attached to each foot, and implement PDR system using dual foot-mounted IMU to reduce the analyzed error. The PDR system using foot-mounted IMU is generally based on an inertial navigation system (INS). To reduce bias and white noise errors, INS is combined with zero velocity update (ZUPT), which assumes that the pedestrian shoe velocity is zero at the stance phase. Although ZUPT could compensate the velocity and position, the heading drift still occurs. When analyzing the characteristics of the position error, the error shows a symmetrical characteristic. In order to reduce this error, the previous researches compensate for both positions by applying feet position constraints. The algorithm consists of applying a conventional PDR system to each foot and fusion algorithm combining both. The PDR system using foot-mounted IMU, one on each foot, is based on integration approach separately. The positions of both feet should be in a circle with a radius as step length during walking. The designed filter is constrained so that the position of both feet are in a circular boundary. The heading error that is symmetrically drifted is corrected by the position constraint when the pedestrian moves straight. Experimental results show the performance and usability of each previous algorithm to compensate for symmetric heading errors.

RNN-Aided Human Velocity Estimation from a Single IMU

“…Chang et al [44] learn intuitive physics, Karl et al [45] create state-space models, and Stewart et al [46] monitor neural networks (NN) via physical knowledge. Others [8,47,48] learn to limit a system drift or to get a more precise zero velocity update (ZUPT) phase to compensate for the errors of inertial systems [11,12,49]. However, most of these approaches focus on the analysis of human movement and do not provide velocity, orientation, or position.…”

“…Konda et al [51] propose a CNN and use data from two synchronized sensors to predict the direction and the velocity changes. Carrera et al [52] use two or more IMUs or a military-grade IMU [12] to reduce accumulating errors and increase the confidence of their algorithms. Others learn intuitive physics [3], design state space models [53], or monitor NNs via physical knowledge [46,54].…”

“…A combination with ZUPT [8,11,60] reduces the sensor drift (error < 0.25 m/s) [8,57] and, hence, provides a higher accuracy. However, at higher movement speeds, ZUPT fails due to noise and movement artifacts [8,11,12,58]. Other methods first classify a movement type with ML or DL methods and then select a suitable step length estimator.…”

“…However, this only works well if both the accelerometer and the gyroscope are mounted on the feet [7,8] or on the spine [3,9,10]. Such methods then do not require additional external sensors or machinery that provide accurate positions over a longer period of time [11,12]. The technical challenges of sensor drift and the decomposition of the acceleration are still unsolved and even rotation-invariant methods that use frequency-based parameters still suffer from poor accuracy [7,10].…”

“…To improve their accuracy dozens of motion models have been examined: models of empirical relationships [15,16], biomechanical models [15,17,18], linear models [19], non-linear models [20][21][22], and models that are based on regression [19,[23][24][25]. However, they also suffer from orientation variance and provide less accuracy than data-driven state-of-the-art techniques [11,12,26].…”