Model-based gait extraction and recognition

Abstract

Extracting full-body motion from monocular video, sequences for gait recognition is an important and difficult problem.  Very often, the motion will be highly articulated and have complex changing boundaries and images may suffer from high level correlated and random noise from the real world.  Moreover, the large variations of the appearances of walking people caused by, for instance, carrying objects or wearing clothing, make the problem even more complicated.  In this thesis, we propose a consistent and easily extensible Bayesian framework for the gait extraction problem using strong prior knowledge.  This knowledge is imposed by a single two-dimensional articulated model having both time-invariant (static) and time-variant (dynamic) parameters.  The model is easily extended to handle the variations of body shapes.  To exploit the dynamics of human walk, we use a hidden Markov model to detect the phases of images in: a walking cycle.  The PDF projection theorem is introduced to learn the observation probability distributions accurately.  We build a strong prior model from the statistics of the parameters of the articulated model, which are learned from noise-free indoor training data.  The system parameters are first bootstrapped from a small amount of data and then refined by the Bayesian updating.  We demonstrate our approach on both high-quality indoor and noisy outdoor video data, as well as high-quality data with synthetic noise and occlusions added, and walkers with rucksacks, skirts and trench coats.

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata