PhD Defense: Ioannis Papavasileiou

Title: An Algorithmic Framework For Gait Analysis and Gait-Based Biometric Authentication Student: Ioannis Papavasileiou Major advisor: Dr. Song Han Associate Advisors: Dr. Jinbo Bi, Dr Wenlong Zhang (ASU) Date & time: Friday 8/3 at 11am Location: ITE 336 Abstract: Gait refers to the locomotion achieved through the movement of human limbs and is fairly unique to an individual due to their specific muscular-skeletal structure. However, conditions that affect the nervous system, such as Parkinson’s Disease (PD) and stroke, can cause significant impairment in cognition, motor skills and gait disorders. Consequently, small or large deviations present in someone’s gait could be attributed to possible underlying neurological disorders or to their unique gait patterns. Motivated by that, in this thesis we examine different types of deviations present in someone’s gait and we design and develop an algorithmic framework that identifies such deviations caused from neurological disorders or unique individual behavior. First, we present two methods for gait analysis. To objectively extract gait phases, an infinite Gaussian mixture model is proposed to classify different gait phases, and a parallel particle filter to estimate and update the model parameters in real-time. To objectively classify gait disorders caused by PD and stroke diseases and to facilitate gait physical therapy, an advanced machine learning method is used, multi-task learning, to jointly train classification models of a subject’s gait. The proposed method highly improves the performance when compared to the baseline solutions and is able to identify parameters that can be used to distinguish between the gait abnormalities and help therapists provide targeted treatment in clinics. Finally, we present a new approach for identifying unique gait patterns, that can be attributed to unique individual behavior, and provide gait-based biometric authentication. Wearable sensors such as smart shoes or socks are used as gait sensing devices and are capable of recording acceleration and ground contact forces. The proposed approach relies on multimodal learning, with a neural network of bimodal-deep auto-encoders. The proposed methodology outperforms existing solutions, and provides robust and user friendly mobile authentication experience.