Robust Vehicle Registration Method Based on 3D Model for Traffic Surveillance Application

Abstract

D-2D vehicle registration provides a new way for vehicle recognition, localization and tracking in traffic surveillance systems. In this paper we present two novel fitness functions to measure 3D-2D vehicle matching, where 3D wire-frame model is used. Unlike the existing vehicle registration methods, we group the model's wireframes into the important and unimportant ones in view of the disaccord between real vehicle and the wire-frame model. The important wireframes generally well fit the corresponding image edges, whereas the image edges corresponding to the unimportant wireframes may not exist due to the streamlined design of real vehicle. For more accurate matching, the fitting of the important wireframes is underlined in both two fitness functions. In the first fitness function, the larger weight coefficient is assigned to the fitting of the important wireframes; in the second fitness function, two different functions are used for the fitting of the model's wireframes instead of two different weight coefficients. Experiments on real traffic videos verify the correctness and robustness of the proposed fitness functions.