Adaptive gait detection based on foot-mounted inertial sensors and multi-sensor fusion

Gait detection plays an important role in areas where spatial-temporal gait parameters are needed. Inertial sensors are now sufficiently small in size and light in weight for collection of human gait data with body sensor networks (BSNs). However, gait detection methods usually rely on careful sensor alignment and a set of rule-based thresholds, which are brittle or difficult to implement. This paper presents an adaptive method for gait detection, which models human gait with a hidden Markov model (HMM), and employs a neural network (NN) to deal with the raw measurements and feed the HMM with classifications. Six gait events are involved for a detailed analysis, i.e., heel strike, foot flat, mid-stance, heel off, toe off, and mid-swing. In order to obtain enough gait data for training a gait model, the gait events are labeled by a rule-based detection method, in which the predefined rules are verified with an optical motion capture system. Experiments were conducted by nine subjects, based on a dual-sensor configuration with one sensor on each foot. Detection performance is quantified using metrics of accuracy, sensitivity and specificity, and the averaged performance values are 98.11%, 94.32% and 98.86% respectively with a timing error less than 2.5 ms.