Two-stage appearance-based re-identification of humans in low-resolution videos

The objective of human re-identification is to recognize a specific individual on different locations and to determine whether an individual has already appeared. This is especially in multi-camera networks with non-overlapping fields of view of interest. However, this is still an unsolved computer vision task due to several challenges, e.g. significant changes of appearance of humans as well as different illumination, camera parameters etc. In addition, for instance, in surveillance scenarios only low-resolution videos are usually available, so that biometric approaches may not be applied. This paper presents a whole-body appearance-based human re-identification approach for low-resolution videos. The method is divided in two stages: first, an appearance model is computed from several images of an individual and pairwise compared to each other. The model is based on means of covariance descriptors determined by spectral clustering techniques. In the second stage, the result is refined by learning the appearance manifolds of the best matches. The proposed approach is tested on a multi-camera data set of a typical surveillance scenario and compared to a color histogram based method.