Person Identity in Gait Patterns: A Spatial-Temporal Approach for Non-Invasive Gait Identification in Realistic Environments Using Deep Machine Learning and Computer Vision

Person identification is a substantial problem of vital importance with critical applications in public security and law enforcement, such as missing person identification and crime suspect identification. However, existing works have severe limitations, for example, many approaches rely upon face identification, which can easily be avoided by hiding or disguising the face. Furthermore, these approaches are limited in real-world environments due to the requirement of unobstructed high-quality images, which are not always available in uncontrolled environments. An alternative method is gait identification, where people are identified by their manner of walking. This is a convenient approach as it may be performed without the knowledge or consent of the subject, it does not require high-quality images, and gait is a difficult biometric to disguise. However, many current approaches are limited and often affected by appearance, recording environment, viewing angle, and occlusion. This thesis addresses the above issues of gait identification by developing a skeleton-based approach, where bodily keypoints, such as limb joints, are estimated and used to extract gait information. For the first time, a novel method of occluded gait identification is introduced, proposing a composite of state estimation and spatiotemporal deep learning to reconstruct missing keypoints. Next, a viewing angle classifier is implemented to identify the appropriate person identification model for each category of viewing angle. The reported results suggest that this approach provides state-of-the-art performance and is not affected by changes in appearance, viewing angle, or environments.