Gait Phase Detection for Lower-Limb Exoskeletons using Foot Motion Data from a Single Inertial Measurement Unit in Hemiparetic Individuals

Abstract

Due to the recent rise in the use of lower-limb exoskeletons as an alternative for gait rehabilitation, gait phase detection has become an increasingly important feature in the control of these devices. In addition, highly functional, low-cost recovery devices are needed in developing countries, since limited budgets are allocated specifically for biomedical advances. To achieve this goal, this paper presents two gait phase partitioning algorithms that use motion data from a single inertial measurement unit (IMU) placed on the foot instep. For these data, sagittal angular velocity and linear acceleration signals were extracted from nine healthy subjects and nine pathological subjects. Pressure patterns from force sensitive resistors (FSR) instrumented on a custom insole were used as reference values. The performance of a threshold-based (TB) algorithm and a hidden Markov model (HMM) based algorithm, trained by means of subject-specific and standardized parameters approaches, were compared during treadmill walking tasks in terms of timing errors and the goodness index. The findings indicate that HMM outperforms TB for this hardware configuration. In addition, the HMM-based classifier trained by an intra-subject approach showed excellent reliability for the evaluation of mean time, i.e., its intra-class correlation coefficient (ICC) was greater than 0 . 75 . In conclusion, the HMM-based method proposed here can be implemented for gait phase recognition, such as to evaluate gait variability in patients and to control robotic orthoses for lower-limb rehabilitation.

Keywords: force sensitive resistors; gait phase detection; hidden Markov model; inertial measurement unit; inertial motion data; standardized parameters training; subject-specific training; threshold-based algorithm.

Figure 1

Flowchart of feature extraction from inertial motion data. This chart shows how a feature list is updated based on the fulfilment of certain conditions. The occurrence of each feature and their corresponding conditions are sequentially assessed in the following manner: Crossed High Threshold (Ref. 1), Crossed Low Threshold (Ref. 2), Crest Middle (Ref. 3), Crest (Ref. 4), Trough Middle (Ref. 5), Trough (Ref. 6), and Neutral (Ref. 7). When a new gait phase is detected, the feature list is emptied for further searches.

Figure 2

Threshold-based gait phase detection by means of an inertial sensing system over two gait cycles. Feature-based decisions are made to identify each gait phase start: HS (first dashed line), FF (second dashed line), HO (third dashed line) and SP (fourth dashed line).

Figure 3

State machine of the threshold-based algorithm. The transition conditions are based on features found in the angular velocity ($G_y$) and linear acceleration ($A_y$) signals.

Figure 4

Possible transitions $\left(a_{ij}\right)$ among four states $\left(S_i\right)$ of a continuous HMM according to a left–right model. Each model state is paired to a gait phase, whose emissions are modeled using a Gaussian mixture model with three components. $a_{ij}$ denotes the transition probability from state $S_i$ to state $S_j$.

Figure 5

Flowchart that illustrates the validation methodology of HMM. A model is trained by means of the Baum–Welch algorithm, after applying feature extraction to the acquired dataset. The optimal state sequence is then computed through the Viterbi algorithm by using feature vectors from the test dataset, and the performance evaluation is conducted with respect to gait phases labels drawn from the reference system.

Figure 6

Experimental setup. Each subject was instrumented on their dominant/affected side with a custom FSR-equipped insole and an IMU placed on the dorsal side of their foot. Foot motion data were captured by using a Raspberry Pi running ROS to ensure synchronized data acquisition.

Figure 7

Gait phase detection based on FSR activation patterns. The gait cycle may be divided into seven phases, but four detected phases are sufficient for control purposes of an AFO. With this particular granularity, gait phases can be sequentially referenced from FSR data as follows: (i) HS State: Heel FSR is activated; (ii) FF State: first and fifth metatarsal FSR must be activated (black dots), and the activation of heel and toe FSRs is considered depending on specific study subject (grey dots); (iii) HO State: toe FSR must be activated, and the activation of first metatarsal FSR is considered depending on specific study subject; and (iv) SP State: all FSRs must be deactivated. Adapted from [21].

Figure 8

(a) Mean time (MT); and (b) Coefficient of variation (CoV) evaluated for stride and each gait phase for the reference system (FSR), and each detection algorithm: threshold-based (TB) and HMM-based algorithm, both using subject-specific training (SST) and standardized parameter training (SPT). The left bar chart corresponds to the results of healthy subjects (H), and the right graph to those of pathological subjects (P). Asterisks denote statistically significant differences (Bonferroni, $p<0.05$).

Figure 9

Goodness index (G) for healthy subjects (H) and subjects with mobility impairments (P) for each detection algorithm: threshold-based (TB), and HMM-based algorithm, using subject-specific training (SST) and standardized parameter training (SPT). Asterisks indicate statistically significant differences among classifiers (Bonferroni, $p<0.05$).