Precise Edge Tracking of Vehicles Using a Static Camera Setup

Abstract

We describe a novel approach to highly accurate model-based vehicle tracking using multiple static cameras. The proposed system is able to cope with complex lighting situations by using a particle filtering framework based on edge features. The correct pose is determined by rendering multiple pose hypotheses of the 3D model and comparing the observation with the rendering. The main contribution of this work is the adaptation of the edge-based particle filter to the domain of vehicle tracking by exploiting the prevalent constraints. The degrees of freedom of vehicle motion can be reduced to three by taking advantage of the ground plane constraint. Furthermore, a vehicle motion model is employed to reduce the number of probable poses, thus increasing accuracy and speed. We exploit graphics processing unit functionalities to achieve real-time operation.