Contour Extraction of Human with Single-Pixel Width

Abstract

A novel algorithm is presented to acquire accurate human contour. For current algorithms extracting contour with multi-pixels width, it is difficulty in obtaining accurate distance between centroid and any point on human contour for gait recognition. For connective human contour, we use candidate regions and vectors to calculate and compare the angles of adjoining vectors, so that we get point set of human contour and describe human contour of single-pixel width. For disjoint contour, image pre-processing is implemented to fuse disjoint silhouettes with their centroids. Subsequently, we go on performing the algorithm of complete contour to get accurate contour. This algorithm solves the problem that any point on human contour locates accurately. It takes the advantages in image processing of human silhouette, particularly in accurate extraction of human contour for gait or posture recognition.