Vehicle shape approximation from motion for visual traffic surveillance

Abstract

In this paper, a vehicle shape approximation method based on the vehicle motion in a typical traffic image sequence is proposed. In the proposed method, instead of using the 2D image data directly, the intrinsic 3D data is estimated in a monocular image sequence. Given the binary vehicle mask and the camera parameters, the vehicle shape is estimated by the four stages shape approximation method. These stages include feature point extraction, feature point motion estimation between two consecutive frames, feature point height estimation from motion vector, and the 3D shape estimation based on the feature point height. We have tested our method using real world traffic image sequence and the vehicle height profile and dimensions are estimated to be reasonably close to the actual dimensions.