Pose invariant gait analysis and reconstruction
One of the unique advantages of human gait is that it can be perceived from a distance. A varied range of research has been undertaken within the field of gait recognition. However, in almost all circumstances subjects have been constrained to walk fronto-parallel to the camera with a single walking speed. In this thesis we show that gait has sufficient properties that allows us to exploit the structure of articulated leg motion within single view sequences, in order to remove the unknown subject pose and reconstruct the underlying gait signature, with no prior knowledge of the camera calibration. Articulated leg motion is approximately planar, since almost all of the perceived motion is contained within a single limb swing plane. The variation of motion out of this plane is subtle and negligible in comparison to this major plane of motion. Subsequently, we can model human motion by employing a cardboard person assumption. A subject's body and leg segments may be represented by repeating spatio-temporal motion patterns within a set of bilaterally symmetric limb planes. The static features of gait are defined as quantities that remain invariant over the full range of walking motions. In total, we have identified nine static features of articulated leg motion, corresponding to the fronto-parallel view of gait, that remain invariant to the differences in the mode of subject motion. These features are hypothetically unique to each individual, thus can be used as suitable parameters for biometric identification. We develop a stratified approach to linear trajectory gait reconstruction that uses the rigid bone lengths of planar articulated leg motion in order to reconstruct the fronto-parallel view of gait. Furthermore, subject motion commonly occurs within a fixed ground plane and is imaged by a static camera. In general, people tend to walk in straight lines with constant velocity. Imaged gait can then be split piecewise into natural segments of linear motion. If two or more sufficiently different imaged trajectories are available then the calibration of the camera can be determined. Subsequently, the total pattern of gait motion can be globally parameterised for all subjects within an image sequence. We present the details of a sparse method that computes the maximum likelihood estimate of this set of parameters, then conclude with a reconstruction error analysis corresponding to an example image sequence of subject motion.