Abstract
New technology in motion sensing has allowed for an advance in gait analysis and cerebral palsy diagnostics. Body sensors networks have emerged as a promising tool for gait analysis and activity recognition, and orthotic treatment is prevalent among those with cerebral palsy. In this work, a framework for activity classification using inertial sensors mounted on ankle foot orthoses (AFOs) is presented. A hybrid decision tree-nearest neighbor algorithm classifies activities and postures using subject-specific training. To evaluate sensitivities, eight volunteer subjects wore modified bilateral AFOs with shank and foot mounted triaxial accelerometers and gyroscopes. The AFOs were fitted with hardware to induce different gait perturbations: free rotation of the ankle, plantarflexion or “equinus” gait, and locked ankle joint. For each condition, the subject performed eight gait activities at varied slopes and standing, sitting, and lying postures.
Using data from the test protocols, the classification framework was performed to assess training data and number of nearest neighbor effects on classification sensitivity. These tests showed high sensitivity levels even with training data which did not outsize the test data, and that the highest sensitivities were obtained using only one nearest neighbor for comparison. Using optimal training data size and one nearest neighbor methods, forced activity classification was performed using the classification framework to assess sensitivity results for each activity. The results from these tests indicated high levels of sensitivity in recognizing predefined gait events for all perturbed conditions, and that semi-natural movement could be classified to some degree using annotated, predefined movement for training. Subsequently, feature selection using cluster analysis was explored, indicating that feature reduction based on significance thresholds improved results for semi-natural activity classification. Finally, a declassification metric was examined, and results showed increased specificity results using declassification. Our results indicate that AFOs are a suitable sensor platform for future research in activity classification and gait monitoring in AFO users with perturbed gait using limited training data.
