Real-time gait detection based on Hidden Markov Model: Is it possible to avoid training procedure?

Abstract

In this paper we present and validate a methodology to avoid the training procedure of a classifier based on an Hidden Markov Model (HMM) for a real-time gait recognition of two or four phases, implemented to control pediatric active orthoses of lower limb. The new methodology consists in the identification of a set of standardized parameters, obtained by a data set of angular velocities of healthy subjects age-matched. Sagittal angular velocities of lower limbs of ten typically developed children (TD) and ten children with hemiplegia (HC) were acquired by means of the tri-axial gyroscope embedded into Magnetic Inertial Measurement Units (MIMU). The actual sequence of gait phases was captured through a set of four foot switches. The experimental protocol consists in two walking tasks on a treadmill set at 1.0 and 1.5 km/h. We used the Goodness (G) as parameter, computed from Receiver Operating Characteristic (ROC) space, to compare the results obtained by the new methodology with the ones obtained by the subject-specific training of HMM via the Baum-Welch Algorithm. Paired-sample t-tests have shown no significant statistically differences between the two procedures when the gait phase detection was performed with the gyroscopes placed on the foot. Conversely, significant differences were found in data gathered by means of gyroscopes placed on shank. Actually, data relative to both groups presented G values in the range of good/optimum classifier (i.e. G ≤ 0.3), with better performance for the two-phase classifier model. In conclusion, the novel methodology here proposed guarantees the possibility to omit the off-line subject-specific training procedure for gait phase detection and it can be easily implemented in the control algorithm of active orthoses.