Human Identification using Silhouette Gait Data

Identificaiton of people using a set of gait image sequences in varing situations is a challenging problem. We use a bilinear model to separate two independent factors, gait style and phase. N-normalized gait poses is defined and generated by embedding gait image sequences to a standard lower dimensional manifold and learning mapping from the manifold to every pixel. This normalized gait phase is used to collect aligned gait poses from different speed walking image sequence. We identify gait style-vectors, which represent factors invariant to gait pose. Gaitcontent vectors, dependent on the environment,are adapted to the new environment. Using a boosted gait content vector, we get a better human identification accuracy than when using the original phase vector before identifying gait content vector. Support vector machines are used to improve classification.