Robust parameter estimation in lane following using a committee of local expert networks

This research proposes a novel sensor fusion system for lane following in autonomous vehicle navigation. The redundant sensors are a camera positioned in front of the rear view mirror of the vehicle and a map matching system consisting of a DGPS and a digital map. A local estimate of the road curvature is obtained with the use of the extended Kalman filter, while the global estimate is obtained from the map matching system. A fuzzy logic "gating network" is used to partition the input space into clusters, each associated with a RBF expert network. Training of the complete system is carried out on-line. Simulation results demonstrate the superior performance of the fusion scheme.