Our lane fusion system filters the output of lane detection to estimate two things:
1) The current centerline we wish follow
2) The bias between our GPS postion and actual location within the RNDF

The current centerline is computed by first merging adjacent lane markings – for example, a double-yellow line. Then, for each lane marking detected, potential centerlines are extrapolated on either side of the marking, and centerlines near each other are merged together. The correct centerline is chosen based on angle to and distance from the car’s current position, the strength of the detected lane, and other measures.

The RNDF provides advance knowledge of the route network layout. To account for bias in our GPS position and heading, we attempt to match lanes detected with our vision system to those indicated in the RNDF, giving us a better estimate of our position on the course. If no lane markings are detected, we can switch to using the RNDF and drive solely based on GPS until lane markings re-appear.

A sample frame from lane fusion is shown below – the yellow and grey curves show the detected yellow and white markings, the green and light blue curve shows the estimated vision centerline, the pink curve shows the non-corrected RNDF centerlines, and the dark red curve shows the corrected RNDF centerlines.

The benefits of our lane fusion system include:
-The ability to drive without an RNDF or GPS when lane markings are visible
-Robustness to GPS drift or bias with respect to RNDF localization
-Extensibility to generic “road markings” such as berms, guard rails, etc.