Model-Free Representations for Gait Recognition

Gait is the manner of walking, and gait recognition concerns the identification of people in video sequences by the way they walk. There is a number of advantages that makes gait valuable as a biometric. For instance, it is possible to detect and measure gait even in low resolution video, where it is often difficult to get other modalities, e.g., face or iris information at high enough resolution for recognition purposes. In addition, gait is difficult to disguise or conceal. Psychophysical studies indicate that humans have the capability for recognising people from even impoverished displays of gait, revealing the presence of identity information in the gait. Henceforth, it is interesting to study the utility of gait as a biometric. The goal of this thesis is to extract the motion information contained in the video sequences of the human gait and to exploit these information in means that facilitate individual recognition. To that end, four model-free methods are proposed. The proliferation of Fourier descriptors in shape analysis inspires the creation of the gait representation incorporating Fourier descriptors. The second is a method that captures the recency of gait using motion history, described by the histograms of oriented gradients. Since gait is a spatiotemporal phenomenon, it is also intuitive to explore the possibilities of characterising these spatiotemporal patterns using temporal motion patterns and statistical distribution. This notion led to the third and fourth methods; the former encodes the transient binary patterns and the latter exploits the statistical mean and variance of the silhouette deformation in the gait cycle