Model driven state estimation for target pursuit

Abstract

Autonomous target pursuit is an extremely useful technology for surveillance applications. In this paper, we derive and evaluate, in a realistic simulation, a novel tracking algorithm for vision-based pursuit. We assume a simple ground-based surveillance robot equipped with a single monocular camera. For the sensor, we propose the use of a color histogram based region tracker. We integrate models of the robot's kinematics and the target's dynamics with a model of the color region tracking sensor via an extended Kalman filter. Detailed simulation results demonstrate that the tracking algorithm substantially reduces the relative position estimation error introduced by noisy color region tracking. The algorithm thus enables target pursuit based on an extremely noisy but simple and low cost sensor.