Omnidirectional Gait Identification by Tilt Normalization and Azimuth View Transformation
Résumé
We propose an effective matching method of gait sequences containing changes of observed tilt and azimuth views provided by an omnidirectional camera. Given gait image sequence, observed tilt view and depth to the subject are normalized by re-projecting the images to a virtual tilt-free image plane orthogonally, and a series of tilt- and size-normalized gait silhouette images are constructed as Gait Silhouette Volume (GSV). On the other hand, a relatively wide range of azimuth view changes can be continuously observed by tracking the subject in the omnidirectional image, multi-view frequency-domain features can be obtain from the GSV as attractive multiple cues for identification. Observed azimuth view ranges of two gait sequences do, however, not always coincide, therefore a View Transformation Model (VTM) supported by reference view addition with geometrical model is exploited to synthesize missing-view gait features for each gallery and probe sequence in matching phase. Experiments of gait identification including 22 subjects from different view ranges demonstrate the effectiveness of the proposed method.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...