Classifying Floor Accelerations for Gait Estimation and Detection

Structural vibrations are traditionally used to learn about structural systems, such as determining the dynamic characteristics of a structure and evaluating its structural performance under certain dynamic loads. Recently, researchers have proposed the use of structural vibrations to know more about the occupants of the structure in addition to the structure itself. One example of such would be to determine when a person falls, or to identify the location of a person walking in a building. Using structural vibrations to study occupants can be directly applicable to many fields, ranging from healthcare to security applications. One powerful use of structural vibrations in healthcare is the estimation of gait parameters such as walking speed and cadence because changes in these gait parameters could provide indications of underlying medical conditions. The objective of this dissertation isto formulate a vibration-based gait classifier and a Bayesian gait speed estimator from acceleration signals of floor vibrations.The proposed classifier can detect gait events and perform gait speed estimations using a Bayesian framework. The first step to identify gait parameters from floor vibrations is to determine if the acceleration records were induced by a person walking or by other events. The identified acceleration records produced by a person walking can then be further processed to estimate gait parameters based on the acceleration induced by single-step events. Traditional classification techniques, like artificial neural networks or support vector machines, require previous training, which is time-consuming. Additionally, a large amount of training samples are required for the most used classification techniques to provide an accurate classification. Arguably, Neural Networks (NN) can be the most intensely research tool used for classification purposes due to the excellent classification capabilities those techniques provides, but the lack of quantitative parameters makes this tool a black-box model due to the large amount of parameters and the typical non-linearity of the activation functions. This dissertation proposes and tests a novel parameter-based classification to extract floor acceleration records produced by a person walking. Data collected using seismic accelerometers in a hallway of an academic institution over several months to validate the proposed algorithms experimentally. Results show that the parameter-based classification can successfully identify if an acceleration record belongs to a person walking with an accuracy of 99.75% over a total of 9569 records. The classifier was also implemented in a real-world scenario inside a senior care facility with 24/7 data acquisition over the course of a year and a half, with a large amount of data collected that is instrumental to society. The proposed Bayesian-based model can perform gait speed estimations based on the characteristic features of a simulated person walking on a straight line.