When establishing correspondence between objects across non-overlapping cameras, the existing methods combine separate likelihoods of appearance and kinematic features in a Bayesian framework, constructing a joint likelihood ...