FINGERPRINTING VEHICLES FOR TRACKING ACROSS NON-OVERLAPPING VIEWS

In this paper, we propose a method to track and identify vehicles across a sparse set of non-overlapping cameras. Iden tification of vehicles across non-overlapping views requir es an algorithm that can address changes in observed images due to both pose and illumination variations. We propose the use of the 3D structure of the vehicle to tackle variation s due to pose, while statistical appearance models account fo r variations due to illumination. We maintain a fingerprint for each vehicle that comprises both the 3-D structure and the appearance model. The proposed algorithm exploits the presence of a ground plane for tracking as well as estimating the 3D structure of the target. The ground plane constraint provides a compact state space representation for tracking . Online recovery of structure provides the capability for ex plicitly addressing changes in pose across cameras. The fingerprintsare shared across the network to generate a watchlist, against which new targets are verified when they enter the sensing field.