Automatic detection of gait events using kinematic data

Abstract

The timing of heel strike (HS) and toe off (TO), the events that mark the transitions between stance and swing phase of gait, is essential when analysing gait. Force plate recordings are routinely used to identify these events. Additional instrumentation, such as force sensitive resistors, can also been used. These approaches, however, include restrictions on the number of steps that can be analyzed and further encumbrance of the subject. We developed an algorithm which automatically determines these times from kinematic data recorded by a motion capture system, which is routinely used in gait analysis laboratories. The foot velocity algorithm (FVA) uses data from the heel and toe markers and identifies features in the vertical velocity of the foot which correspond to the gait events. We verified the performance of the FVA using a large data set of 54 normal children that contained both force plate recordings and kinematic data and found errors of (mean ± standard deviation) 16 ± 15 ms for HS and 9 ± 15 ms for TO. The algorithm also worked well when tested on a small number of children with spastic diplegia. We compared the performance of the FVA with another kinematic method previously described. Our foot velocity algorithm offered more accurate results and was easier to implement than the previously described one, and should be applicable in a variety of gait analysis settings.

Introduction

The definition of individual gait events is an essential starting point for nearly all aspects of gait analysis. By convention, initial foot contact is generally taken as the starting point of a complete gait cycle and marks the beginning of the stance phase. Termination of foot contact marks the beginning of the swing phase. In normal gait these events correspond to heel strike (HS) and toe off (TO), respectively. The timing information of these events is used in the analysis of temporal gait parameters such as stride time and periods of single and double support, and allows for the time normalization of data per gait cycle. This analysis is necessary for examining ensemble averages of kinematic, kinetic and EMG patterns over a number of gait cycles, and facilitates comparison between different subjects and conditions.

The gold standard method of defining gait events is based on the use of a force plate. A simple threshold on the force level can accurately provide the timing of HS and TO, provided a “clean” force plate hit has occurred, one in which a single foot lands entirely on the force plate. This method is generally restricted to a gait laboratory setting and the number of available force plates (often one or two) limits the number of consecutive gait cycles that can be analyzed. Researchers have developed alternative systems to overcome these restrictions such as instrumented walkways [1], [2], using force sensitive resistors attached to the feet [3], [4] or wearing specially instrumented shoes [5]. Timing information can also be extracted from the signals recorded from additional sensors such as accelerometers [6], [7] or gyroscopes [8].

Motion analysis systems are often used to record kinematic data in gait analysis, based on trajectories of markers attached to landmark positions on the subject. Both experienced and untrained raters, following specific instructions, can generally identify HS and TO to within acceptable tolerances manually, by examining the marker trajectories and velocity plots [9], [10]. Automatic algorithms which use thresholds on the height or velocity of markers have been used with more variable results [10], [11]. Stanhope et al. [12] developed a method which accurately determined timing information using a kinematic model. However, this model was individual-specific and required a separate method, such as the use of a force plate, to determine the initial occurrence of gait events.

Hreljac and Marshall [13] developed algorithms for determining HS and TO based on the displacement, acceleration and jerk (the derivative of acceleration) of heel and toe markers. Their algorithms determined the location of characteristic peaks, troughs and zero crossings to determine HS and TO times. However, the location of the points had to be estimated first to ensure the correct points were identified. Their method also used optimal filtering [14] of each marker as an initial step. The results of such an algorithm were shown to be highly sensitive to the choice of cutoff frequency [15]. Other researchers have shown that various methods for determining the optimal cutoff frequency produce different results and have proposed the selection of different cutoff frequencies when higher derivatives of the displacement time data are to be calculated [16].

We developed a new algorithm for determining HS and TO that is based on a simple velocity curve derived from heel and toe marker trajectories, filtered with a single cutoff frequency. We validated the algorithm by comparing it to the timing of the same gait events as determined with the use of a force plate, for both normal gait and a limited number of patients with spastic diplegia.