On optimal camera parameter selection in kalman filter based object tracking

Abstract In this paper we present an information theoretic framework that pro-vides an optimality criterion for the selection of the best sensor data regarding state estimation of dynamic system. One relevant application in practice is track-ing a moving object in 3–D using multiple sensors. Our approach extends previ-ous and similar work in the area of active object recognition, i.e. state estimation of static systems. We derive a theoretically well founded metric based on the con-ditional entropy that is also close to intuition: select those camera parameters that result in sensor data containing most information for the following state estima-tion. In the case of state estimation with a non–linear Kalman filter we show how that metric can be evaluated in closed form. The results of real–time experiments prove the benefits of our general approach in the case of active focal length adaption compared to fixed focal lengths. The main impact of the work consists in a uniform probabilistic description of sensor data selection, processing and fusion.