Recognizing walking people

Abstract

We present a method for recognition of walking people in monocular image sequences based on extraction of coordinates of specific point locations on the body. The method works by comparison of sequences of recorded coordinates with a library of sequences from different individuals. The comparison is based on the evaluation of view in-variant and calibration independent view consistency constraints. These constraints are functions of corresponding image coordinates in two views and are satisfied whenever the two views are projected from the same 3D object. By evaluating the view consistency constraints for each pair of frames in a sequence of a walking person and a stored sequence we get a matrix of consistency values that ideally are zero whenever the pair of images depict the same 3Dposture. The method is virtually parameter free and computes a consistency residual between a pair of sequences that can be used as a distance for clustering and classification. Using interactively extracted data we present experimental results that are superior to those of previously published algorithms both in terms of performance and generality.