Stride History Image: A New Feature Representation for Pedestrian Identification

Abstract

We describe representations of gait appearance features for the purpose of pedestrian identification. We construct two kinds of scalar valued spatiotemporal templates in 2D images, backward stride history image (bSHI) and forward stride history image (fSHI), to represent how change of pedestrian silhouettes is evolved. To define the stride length and ensure an invariant analysis due to differences in gait period, we perform a gait period estimation procedure using nonlinear analysis beforehand. We make use of a family of multi-layer windows to capture SHI image signature in the form of combined moment feature vectors for individual identification. The proposed approach achieves identification capability by an 86.56% CCR on Soton database and improvements are seen with respect to other methods.