Multiple vehicle 3D tracking using an unscented Kalman filter

Abstract

This article describes a system to track vehicles on images taken from a mobile platform. The objective is to determine the position and velocity of vehicles ahead of the mobile platform, in order to make possible the prediction of their position in future instants of time. This problem is addressed by modeling a 3D dynamic system, where both the acquisition platform and the tracked vehicles are represented in a state vector. From measurements obtained in every frame, this state vector is re-estimated using an unscented Kalman filter, instead of the extended Kalman filter used in previous works. Assuming that vehicles progress on a flat surface, a novel model of their dynamics is proposed, which explicitly considers constraints on the velocity. With respect to previous approaches, this model improves tracking reliability, since the estimation of unfeasible states is avoided. Experiments on real sequences display promising results, although a more systematic evaluation of the system should be done.