Abstract:
When establishing object correspondence across
non-overlapping cameras, the existing methods combine separate
likelihoods of appearance and kinematic features in a Bayesian
framework, constructing a joint likelihood to compute the probability
of re-detection. A drawback of these methods is not having
a proper approach to reduce the search space when localizing
an object in a subsequent camera once the kinematic and
appearance features are extracted in the current camera. In this
work we introduce a novel methodology to condition the location
of an object on its appearance and time, without assuming
independence between appearance and kinematic features, in
contrast to existing work. We characterize the linear movement
of objects in the unobserved region with an additive Gaussian
noise model. Assuming that the cameras are affine, we transform
the noise model onto the image plane of subsequent cameras. We
have tested our method with toy car experiments and real-world
camera setups and found that the proposed noise model acts as
a prior to improving re-detection. It constrains the search space
in a subsequent camera, greatly improving the computational
efficiency. Our method also has the potential to distinguish
between objects similar in appearance, and recover correct labels
when they move across cameras.
