An investigation of the impact of gait segmentation on accelerometry-based inclined terrain classification

Traditional methods of energy expenditure estimation, in the free-living environment, attempted using accelerometry operate without knowledge of the slope of the terrain which is being traversed. The ability to recognise the gradient of the walking surface will most likely improve upon these simplistic energy estimates. This paper expands upon previous work in this area, and investigates the benefit of step-by-step segmentation of the accelerometry signal in classifying the various gradients. Tri-axial accelerometry signals from 12 subjects, performing 30 s of walking on 4 different gradients (up and down paved ramps with gradients of 4.8% and 17.2%), were collected. A feature subset selection search procedure was applied to find the optimal subset of 65 extracted features which maximise the classification accuracy, performed with a Gaussian Mixture Model (GMM) classifier, as estimated using six-fold cross-validation. An overall classification accuracy of 94.83% was achieved using 13 features, for the four-class problem. There was an improvement of 4.1% upon the same classification task, without knowledge of the start/end times of individual steps, indicating that segmentation of the accelerometry signals at a step-by-step resolution is important for the automated classification of terrain gradient during walking.