High-Knee-Flexion Posture Recognition Using Multi-Dimensional Dynamic Time Warping on Inertial Sensor Data

Abstract

Relating continuously collected inertial data to the activities or postures performed by the sensor wearer requires pattern recognition or machine-learning-based algorithms, accounting for the temporal and scale variability present in human movements. The objective of this study was to develop a sensor-based framework for the detection and measurement of high-flexion postures frequently adopted in occupational settings. IMU-based joint angle estimates for the ankle, knee, and hip were time and scale normalized prior to being input to a multi-dimensional Dynamic Time Warping (mDTW) distance-based Nearest Neighbour algorithm for the identification of twelve postures. Data from 50 participants were divided to develop and evaluate the mDTW model. Overall accuracies of 82.3% and 55.6% were reached when classifying movements from the testing and validation datasets, respectively, which increased to 86% and 74.6% when adjusting for imbalances between classification groups. The highest misclassification rates occurred between flatfoot squatting, heels-up squatting, and stooping, while the model was incapable of identifying sequences of walking based on a single stride template. The developed mDTW model proved robust in identifying high-flexion postures performed by participants both included and precluded from algorithm development, indicating its strong potential for the quantitative measurement of postural adoption in real-world settings.

Keywords: accelerometer; dynamic time warping; gyroscope; high-knee flexion; inertial sensors; knee osteoarthritis; occupational ergonomics; posture classification.

Figure 1

Depiction of the 13 occupational-inspired motions performed in this study in their fully flexed position. Note that SS and SL are displayed to the left, however were additionally performed to the participant’s right.

Figure 2

Signal processing workflow for the mDTW kNN algorithm. The DTW distances for each joint were calculated between a single normalized trial (referred to as a movement sequence) and the normalized templates for the 12 movement classes included in the developed model. A smaller distance between the sequence and template would represent greater similarity between waveforms. Once each movement sequence had been compared to all templates for the corresponding joint and leg, the DTW distances were combined across joints using custom weighting factors such that the movement class would ultimately be determined based on the mDTW distance.

Figure 3

A representative unnormalized movement sequence S_RKnee for the flexion angle of the right knee during a heels-up squatting (HS) motion trial (A) along with the corresponding heels-up squat template T_RKnee generated from this sequence for the right knee (B).

Figure 4

Templates generated for heels-up squatting based on the right knee across all trials and participants, with mean and standard deviation curves overlayed in red.

Figure 5

Representative warping paths ix and iy derived during the calculation of D_RKnee (C) based on only the right knee between the movement sequence S_RKnee (A) and the template T_RKnee (B) for a heels-up squatting motion trial. Each signal is warped such that the highest level of similarity between waveforms can be achieved. Note that in warping the signals, the length of the final sample may be longer than either original waveform. The dotted and solid lines represent the warped recreations of the movement sequence and template, respectively.