A Segmentation Method of Catadioptric Images for Gait Recognition in Unconstrained Environment

Gait is an emerging biometric technology. It enables biometric at a distance. The first step in gait recognition is the silhouette extraction. However, most of the work involves indoor controlled environment or well-exposed outdoor scenes. Furthermore, they are all applied to perspective-like pictures. This paper addresses a method for silhouette extraction on catadioptric images in indoor and uncontrolled lighting environments. We introduce a new segmentation method based on the K-means clustering algorithm. This method is robust to the stroboscopic effect induced by the light source. We finally present a local method to obtain perspective-like pictures enabling further processing. Principal Component Analysis (PCA) is usually used for dimensionality reduction of datasets. Most of the time, the geometrically asset of the PCA is unused. In this work, we take advantage of this particular point to propose a local unwrapping technique of catadioptric pictures.