Recognizing the reappearance of an object provides a useful functionality f
or automated scene interpretation. While many modalities will in general co
ntribute to this task, this paper focuses on the use of object color proper
ties. Re-identifying objects between cameras requires a calibration of the
respective color spaces, which are a function of both camera and lighting p
roperties. Ways in which such calibration may be accomplished, on-line and
without specialist equipment, are investigated. Matching against previously
observed objects is an image retrieval problem, for which we develop an ex
plicit representation of the observed color distribution. Color histograms
can be built and compared in real time without specialist hardware: we inve
stigate effective representations for classification of object identity. Th
e camera capture noise characteristics are used to define optimal histogram
quantization intervals. A model of the object color properties is built us
ing multiple observations of the same object, acquired with the use of a sp
atial object tracker. Comparative results on real data are presented for si
ngle and multi-camera re-identification, using algorithms which may be exec
uted in real time. (C) 2001 Academic Press.