Pattern recognition from raw spatio-temporal data for gait analysis in healthcare and security

  • Omar Costilla Reyes

Student thesis: Phd

Abstract

Nature has phenomena that can be modelled as a periodic pattern in time and space. Recent advancements in machine learning research and sensor systems offer potential to understand these phenomena with capabilities never attained before. An important spatio-temporal pattern is human motion since it can be a powerful indicator of human well-being and identity. In particular, human gait and footsteps offer a unique human motion pattern. In this work, human motion was investigated in two themes: security and healthcare. In the healthcare theme, the effects of cognitively demanding tasks on human gait patterns were studied to advance research on early detection of neurodegenerative diseases by finding the relationship between a cognitive load and its effects in gait patterns. For the security theme, human footsteps were studied as a unique behavioural pattern for biometric verification systems. The methodology to solve these problems was based on studying spatio-temporal gait and footsteps signals in advanced computational models based on deep machine learning. The models delivered an advanced artificial intelligence that distinguished fine-grained pattern variability in the spatiotemporal signals studied. The data was obtained from two special types of floor sensor systems. Sixty-nine participants were recruited for the healthcare theme while 127 were recruited for the security theme that provided statistically significant results. Visualization methodologies were proposed to investigate the automatic feature learning process of deep learning models. In the healthcare theme, the methodology delivered a top classification F-score of 97.33% to detect effects of cognitively demanding activities on walking patterns. While in the security domain, the methodology delivered an optimal equal error rate of 0.7% in a biometric verification system based solely on footsteps. To justify the optimal performance of the deep learning models, they were compared to non-deep learning methodologies based on classical machine learning techniques. The deep learning methodology outperformed the classical approach by a large margin of 49.56% F-score in the optimal case. The methodologies presented in this work have delivered state-of-the-art performance in both themes and can be applied to other spatio-temporal problems that have similar fine-grained patterns to be identified, For example in climate science, social sciences, neuroscience, epidemiology and transportation.
Date of Award1 Aug 2018
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorKrikor Ozanyan (Supervisor) & Patricia Scully (Supervisor)

Keywords

  • Deep learning
  • Artificial intelligence
  • Biometrics
  • Dual-tasks
  • gait analysis

Cite this

'