Using Unscented Kalman Filter for Road Tracing from Satellite Images

Abstract

The extended Kalman filter with profile matching has been employed to extract road maps in satellite images. This algorithm suffers from several drawbacks that result in its poor performance in difficult situations. To improve its performance in those situations, like junctions and varying road characteristics, we propose to use the unscented Kalman filter which can deal better with the nonlinearity of the road model. Additionally, we use an approach to dissociate the system measurements from the current state prediction of the Kalman filter. This method removes the potential for the instability of the algorithm. Finally, we introduce a technique based on clustering of the road profiles to properly maintain a database on various road characteristics along the way. This way we provide a means to continue tracking even when the road profile undergoes significant variations.