Learning architecture for the recognition of walking and prediction of gait period using wearable sensors

Abstract This work presents a novel learning architecture for the recognition and prediction of walking activity and gait period, respectively, using wearable sensors. This approach is composed of a Convolutional Neural Network (CNN), a Predicted Information Gain (PIG) module and an adaptive combination of information sources. The CNN provides the recognition of walking and gait periods. This information is used by the proposed PIG method to estimate the next most probable gait period along the gait cycle. The outputs from the CNN and PIG modules are combined by a proposed adaptive process, which relies on data from the source that shows to be more reliable. This adaptive combination ensures that the learning architecture provides accurate recognition and prediction of walking activity and gait periods over time. The learning architecture uses data from an array of three inertial measurement units attached to the lower limbs of individuals. The validation of this work is performed by the recognition of level-ground walking, ramp ascent and ramp descent, and the prediction of gait periods. The recognition of walking activity and gait period is 100% and 98.63%, respectively, when the CNN model is employed alone. The recognition of gait periods achieves a 99.9% accuracy, when the PIG method and adaptive combination are also used. These results demonstrate the benefit of having a system capable of predicting or anticipating the next information or event over time. Overall, the learning architecture offers an alternative approach for accurate activity recognition, which is essential for the development of wearable robots capable of reliably and safely assisting humans in activities of daily living.