Target tracking under out of field of view situation using geographic information estimated from background

Abstract

A problem of failure in tracking under out of field of view situation has been considered. We propose a method being able to continue the tracking by taking into account geographic information estimated from the background. We deal with a scene that Unmanned Aerial Vehicle is tracking a car. The tracking sequentially estimates the velocity and posture of the car from an image sequence using particle filter. The geographic information is road. The road area is also estimated using particle filter. Unknown road are extrapolated from known road with cubic splines. Target disappearing from the frame is predicted by state space model which imposes prior knowledge that the car normally moves on the road. Effectiveness of proposed method has been shown through tracking experiments with simulation image sequences.