A robust walking detection algorithm using a single foot-worn inertial sensor: validation in real-life settings

Abstract

Walking activity and gait parameters are considered among the most relevant mobility-related parameters. Currently, gait assessments have been mainly analyzed in laboratory or hospital settings, which only partially reflect usual performance (i.e., real world behavior). In this study, we aim to validate a robust walking detection algorithm using a single foot-worn inertial measurement unit (IMU) in real-life settings. We used a challenging dataset including 18 individuals performing free-living activities. A multi-sensor wearable system including pressure insoles, multiple IMUs, and infrared distance sensors (INDIP) was used as reference. Accurate walking detection was obtained, with sensitivity and specificity of 98 and 91% respectively. As robust walking detection is needed for ambulatory monitoring to complete the processing pipeline from raw recorded data to walking/mobility outcomes, a validated algorithm would pave the way for assessing patient performance and gait quality in real-world conditions.

Keywords: Adaptive threshold; Continuous wavelet transform; Foot-worn sensor; Real-world; Walking detection.

Fig. 1

Multi-sensor wearable system (INDIP); (a) INDIP system attached on the shoe, and (b) picture of two INDIP sensors (right and left feet), which integrate force sensitive resistor pressure insoles, inertial modules (IMUs), and infrared distance sensors

Fig. 2

Flowchart of the walking bouts detection algorithm for one-foot IMU. ωN: raw angular velocity norm; aN: raw acceleration norm; LPF: low-pass filter; CWT: continuous wavelet transform; ${\mathit{TH}}_{\mathit{fixed}}$: fixed threshold (${\mathit{TH}}_{\mathit{fixed}}=100deg/s$ or ${\mathit{TH}}_{\mathit{fixed}}=0.5g$); ${\mathit{TH}}_{\mathit{adapt}}$: adaptive threshold based on the percentile of the obtained amplitude distribution of peaks detected above the fixed threshold (${\mathit{TH}}_{\mathit{fixed}})$; ${\mathit{TH}}_{\mathit{hilbert}}$: Hilbert threshold based on the Hilbert envelope and percentile of the obtained amplitude distribution of peaks detected; ${\mathit{tp}}_i$: time of peak occurrence; StartLoc i: start of the walking bout i; EndLoc i: end of the walking bout i

Fig. 3

Angular velocity signals recorded with the 3D gyroscope on one foot during several gait cycles for one subject. The top panel shows the raw angular velocity signals around the three axes (roll, pitch, yaw). The middle panel shows the raw angular velocity norm (ωN, magenta), the signal after detrending and LPF (ωN-LPF, red), and the signal after continuous wavelet transform (ωN- LPF-CWT, yellow). The bottom panel shows the pitch angular velocity, the continuous wavelet transform (ωN- LPF-CWT, yellow), and the detected strides. Strides are identified as maxima corresponding to mid- swing events (blue circle) with an amplitude higher than a certain threshold (${\mathit{TH}}_{\mathit{fixed}}$, ${\mathit{TH}}_{\mathit{adapt}}$, or ${\mathit{TH}}_{\mathit{hilbert}}$)

Fig. 4

Thresholding methods for peak selection, example based on data from one subject: (a) the filtered signal ωN-LPF-CWT (orange) is shown, as well as the obtained ${\mathit{TH}}_{\mathit{adapt}}$ (dark blue) and the selected peaks (dark blue starts with amplitudes above the thresholds); (b) the Hilbert envelope method is shown with the preselected walking periods (dashed line), and the selected peaks (dark purple dots); (c) amplitude distributions of the peaks, and the thresholds obtained for each of the three tested methods (${\mathit{TH}}_{\mathit{fixed}}$ (green), ${\mathit{TH}}_{\mathit{adapt}}$ (blue) and ${\mathit{TH}}_{\mathit{hilbert}}$ (purple)). The histograms of the peaks obtained for the fixed and adaptive thresholds are identical. In fact, ${\mathit{TH}}_{\mathit{adapt}}$ is defined as the ${10}^{\mathit{th}}$ percentile of the distribution obtained after applying the fixed-threshold ${\mathit{TH}}_{\mathit{fixed}}={100}^{\circ }/s$

Fig. 5

Illustration of walking detection using the different thresholding methods on data recorded in one subject: fixed threshold (green), adaptive threshold (blue) and Hilbert threshold (purple). The reference classification, obtained from the plantar pressure of the INDIP system, is displayed in black at the top of the figure

Fig. 6

ROC curves for performance evaluation as a function of percentile values from 1 to 50% when the angular velocity norm ωN is used as input. The curves are obtained by averaging the results over the 10 subjects when one IMU (continuous line) or two IMUs (dashed line) are used; (a) Adaptive threshold method (${\mathit{TH}}_{\mathit{adapt}}$) based on the percentile of the obtained peak amplitude distribution detected above the fixed threshold $({\mathit{TH}}_{\mathit{fixed}}=100{(}^{\circ }/s))$; (b) Hilbert method: The threshold (${\mathit{TH}}_{\mathit{hilbert}}$) is defined as the percentile of the peak amplitude distribution in the pre-selected walking bouts. The horizontal and vertical dashed lines correspond to a 95% true positive rate and a 4% false positive rate, respectively