Gait Pattern Prediction and Gait-Based Human Identification Using Machine Learning Methods

Abstract

This thesis is concerned with human gait motion learning and prediction,<br/> focusing on differences in gait motion between individuals to address the<br/> problems of personalized gait pattern prediction and gait based biometric<br/> identification.<br/> <br/> We first present an algorithm that predicts individual gait patterns from<br/> a human's anthropometric characteristics, e.g., age, height, weight, gender,<br/> and other related anthropometric data. We adopt the Gaussian process<br/> dynamical model~(GPDM) framework to address the high dimensionality<br/> of human gait motions, and to learn a common stochastic dynamics model<br/> for gaits. We also utilize Gaussian process regression~(GPR) to learn<br/> a mapping from the space of body features to the motion parameters in<br/> the GPDM framework. Using our framework, an entire cycle of individualized<br/> gait motions at arbitrary walking speeds can be predicted from body feature<br/> data.<br/> <br/> We also propose a gait-based identification algorithm based on deep neural<br/> networks. To allow for new subjects whose information is not included in the<br/> original database, we develop a novel autoencoder architecture that is<br/> designed to extract features efficiently from the measured gait motions.<br/> We use the reconstruction error of the trained autoencoder as a similarity<br/> measure between the learned motion and the input, for which user-specified thresholds can<br/> be used for the identification. Our proposed method is able to identify both<br/> subjects in the database as well as unknown subjects, only using the observed<br/> gait motion. Furthermore, it is possible to update the database incrementally<br/> for newly given gait motions.<br/> <br/> Finally, we propose an algorithm used to predict a subject's age, height, weight, and<br/> gender from the subject's gait motion. We construct a neural network architecture<br/> for this purpose by modifying the previous autoencoder structure used for the identification, and<br/> devise an efficient learning strategy utilizing transfer learning to reduce the<br/> prediction error.<br/> <br/> To validate proposed algorithms, we collect gait motions and anthropometric data<br/> from more than a hundred subjects. For our gait motion prediction algorithm, an individualized<br/> gait pattern is generated for observed anthropometric data and at arbitrary walking speeds.<br/> The algorithm also shows 30% less gait pattern prediction errors compared to a frame-by-frame<br/> statistical regression method. In the case of gait identification, our algorithm<br/> shows identification accuracy of 98.5-100%, outperforming the results reported in recent studies.