Switching models for vision-based on-board road detection