Switching Models for Vision-Based On-Board Road Detection